perf: optimize ggml_ext_chunk (#1084 )

feat: add support for custom scheduler (#694 )
--------- Co-authored-by: leejet <leejet714@gmail.com>
2026-10-03 10:48:48 +08:00 · 2025-12-14 01:22:41 +08:00 · 2025-12-13 16:20:02 +08:00 · 2025-12-13 14:27:47 +08:00 · 2025-12-13 14:22:32 +08:00 · 2025-12-13 13:53:21 +08:00
56 changed files with 16726 additions and 1830 deletions
@@ -163,7 +163,7 @@ jobs:
          - build: "avx512"
            defines: "-DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
          - build: "cuda12"
-            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES=90;89;86;80;75"
+            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES='61;70;75;80;86;89;90;100;120'"
          - build: 'vulkan'
            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
    steps:
@@ -176,9 +176,9 @@ jobs:
      - name: Install cuda-toolkit
        id: cuda-toolkit
        if: ${{ matrix.build == 'cuda12' }}
-        uses: Jimver/cuda-toolkit@v0.2.19
+        uses: Jimver/cuda-toolkit@v0.2.22
        with:
-          cuda: "12.6.2"
+          cuda: "12.8.1"
          method: "network"
          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'

@@ -87,6 +87,38 @@ file(GLOB SD_LIB_SOURCES
    "*.hpp"
 )

+find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
+if(GIT_EXE)
+    execute_process(COMMAND ${GIT_EXE} describe --tags --abbrev=7 --dirty=+
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE SDCPP_BUILD_VERSION
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
+    )
+    execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE SDCPP_BUILD_COMMIT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
+    )
+endif()
+
+if(NOT SDCPP_BUILD_VERSION)
+    set(SDCPP_BUILD_VERSION unknown)
+endif()
+message(STATUS "stable-diffusion.cpp version ${SDCPP_BUILD_VERSION}")
+
+if(NOT SDCPP_BUILD_COMMIT)
+    set(SDCPP_BUILD_COMMIT unknown)
+endif()
+message(STATUS "stable-diffusion.cpp commit ${SDCPP_BUILD_COMMIT}")
+
+set_property(
+  SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp
+  APPEND PROPERTY COMPILE_DEFINITIONS
+  SDCPP_BUILD_COMMIT=${SDCPP_BUILD_COMMIT} SDCPP_BUILD_VERSION=${SDCPP_BUILD_VERSION}
+)
+
 if(SD_BUILD_SHARED_LIBS)
    message("-- Build shared library")
    message(${SD_LIB_SOURCES})
@@ -17,6 +17,6 @@ RUN apt-get update && \
    apt-get install --yes --no-install-recommends libgomp1 && \
    apt-get clean

-COPY --from=build /sd.cpp/build/bin/sd /sd
+COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli

-ENTRYPOINT [ "/sd" ]
+ENTRYPOINT [ "/sd-cli" ]
@@ -18,6 +18,6 @@ RUN mkdir build && cd build && \

 FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64 as runtime

-COPY --from=build /sd.cpp/build/bin/sd /sd
+COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli

-ENTRYPOINT [ "/sd" ]
+ENTRYPOINT [ "/sd-cli" ]
@@ -14,6 +14,6 @@ RUN mkdir build && cd build && \

 FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS runtime

-COPY --from=build /sd.cpp/build/bin/sd /sd
+COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli

-ENTRYPOINT [ "/sd" ]
+ENTRYPOINT [ "/sd-cli" ]
@@ -1,5 +1,5 @@
 <p align="center">
-  <img src="./assets/cat_with_sd_cpp_42.png" width="360x">
+  <img src="./assets/logo.png" width="360x">
 </p>

 # stable-diffusion.cpp
@@ -49,6 +49,7 @@ API and command-line option may change frequently.***
    - [Chroma1-Radiance](./docs/chroma_radiance.md)
    - [Qwen Image](./docs/qwen_image.md)
    - [Z-Image](./docs/z_image.md)
+    - [Ovis-Image](./docs/ovis_image.md)
  - Image Edit Models
    - [FLUX.1-Kontext-dev](./docs/kontext.md)
    - [Qwen Image Edit/Qwen Image Edit 2509](./docs/qwen_image_edit.md)
@@ -104,7 +105,7 @@ API and command-line option may change frequently.***
 ### Download model weights

 - download weights(.ckpt or .safetensors or .gguf). For example
-    - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
+    - Stable Diffusion v1.5 from https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 

    ```sh
    curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
@@ -113,7 +114,7 @@ API and command-line option may change frequently.***
 ### Generate an image with just one command

 ```sh
-./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
+./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
 ```

 ***For detailed command-line arguments, check out [cli doc](./examples/cli/README.md).***
@@ -134,6 +135,7 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [🔥Qwen Image Edit/Qwen Image Edit 2509](./docs/qwen_image_edit.md)
 - [🔥Wan2.1/Wan2.2](./docs/wan.md)
 - [🔥Z-Image](./docs/z_image.md)
+- [Ovis-Image](./docs/ovis_image.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
@@ -3,34 +3,10 @@

 #include "ggml_extend.hpp"
 #include "model.h"
+#include "tokenize_util.h"

 /*================================================== CLIPTokenizer ===================================================*/

-__STATIC_INLINE__ std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remove_lora(std::string text) {
-    std::regex re("<lora:([^:]+):([^>]+)>");
-    std::smatch matches;
-    std::unordered_map<std::string, float> filename2multiplier;
-
-    while (std::regex_search(text, matches, re)) {
-        std::string filename = matches[1].str();
-        float multiplier     = std::stof(matches[2].str());
-
-        text = std::regex_replace(text, re, "", std::regex_constants::format_first_only);
-
-        if (multiplier == 0.f) {
-            continue;
-        }
-
-        if (filename2multiplier.find(filename) == filename2multiplier.end()) {
-            filename2multiplier[filename] = multiplier;
-        } else {
-            filename2multiplier[filename] += multiplier;
-        }
-    }
-
-    return std::make_pair(filename2multiplier, text);
-}
-
 __STATIC_INLINE__ std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
    std::vector<std::pair<int, std::u32string>> byte_unicode_pairs;
    std::set<int> byte_set;
@@ -72,6 +48,8 @@ private:
    int encoder_len;
    int bpe_len;

+    std::vector<std::string> special_tokens;
+
 public:
    const std::string UNK_TOKEN = "<|endoftext|>";
    const std::string BOS_TOKEN = "<|startoftext|>";
@@ -117,6 +95,15 @@ private:
        return pairs;
    }

+    bool is_special_token(const std::string& token) {
+        for (auto& special_token : special_tokens) {
+            if (special_token == token) {
+                return true;
+            }
+        }
+        return false;
+    }
+
 public:
    CLIPTokenizer(int pad_token_id = 49407, const std::string& merges_utf8_str = "")
        : PAD_TOKEN_ID(pad_token_id) {
@@ -125,6 +112,8 @@ public:
        } else {
            load_from_merges(ModelLoader::load_merges());
        }
+        add_special_token("<|startoftext|>");
+        add_special_token("<|endoftext|>");
    }

    void load_from_merges(const std::string& merges_utf8_str) {
@@ -201,6 +190,10 @@ public:
        }
    }

+    void add_special_token(const std::string& token) {
+        special_tokens.push_back(token);
+    }
+
    std::u32string bpe(const std::u32string& token) {
        std::vector<std::u32string> word;

@@ -379,25 +372,54 @@ public:
        return trim(text);
    }

+    std::vector<std::string> token_split(const std::string& text) {
+        std::regex pat(R"('s|'t|'re|'ve|'m|'ll|'d|[[:alpha:]]+|[[:digit:]]|[^[:space:][:alpha:][:digit:]]+)",
+                       std::regex::icase);
+        std::sregex_iterator iter(text.begin(), text.end(), pat);
+        std::sregex_iterator end;
+
+        std::vector<std::string> result;
+        for (; iter != end; ++iter) {
+            result.emplace_back(iter->str());
+        }
+
+        return result;
+    }
+
    std::vector<int> encode(std::string text, on_new_token_cb_t on_new_token_cb) {
        std::string original_text = text;
        std::vector<int32_t> bpe_tokens;
        text = whitespace_clean(text);
        std::transform(text.begin(), text.end(), text.begin(), [](unsigned char c) { return std::tolower(c); });

-        std::regex pat(R"(<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[[:alpha:]]+|[[:digit:]]|[^[:space:][:alpha:][:digit:]]+)",
-                       std::regex::icase);
-
-        std::smatch matches;
        std::string str = text;
        std::vector<std::string> token_strs;
-        while (std::regex_search(str, matches, pat)) {
-            bool skip = on_new_token_cb(str, bpe_tokens);
-            if (skip) {
+
+        auto splited_texts = split_with_special_tokens(text, special_tokens);
+
+        for (auto& splited_text : splited_texts) {
+            LOG_DEBUG("token %s", splited_text.c_str());
+            if (is_special_token(splited_text)) {
+                LOG_DEBUG("special %s", splited_text.c_str());
+                bool skip = on_new_token_cb(splited_text, bpe_tokens);
+                if (skip) {
+                    token_strs.push_back(splited_text);
+                    continue;
+                }
                continue;
            }
-            for (auto& token : matches) {
-                std::string token_str = token.str();
+
+            auto tokens = token_split(splited_text);
+            for (auto& token : tokens) {
+                if (on_new_token_cb != nullptr) {
+                    bool skip = on_new_token_cb(token, bpe_tokens);
+                    if (skip) {
+                        token_strs.push_back(token);
+                        continue;
+                    }
+                }
+
+                std::string token_str = token;
                std::u32string utf32_token;
                for (int i = 0; i < token_str.length(); i++) {
                    unsigned char b = token_str[i];
@@ -417,14 +439,13 @@ public:
                bpe_tokens.push_back(encoder[bpe_str]);
                token_strs.push_back(utf32_to_utf8(bpe_str));
            }
-            str = matches.suffix();
        }
-        std::stringstream ss;
-        ss << "[";
-        for (auto token : token_strs) {
-            ss << "\"" << token << "\", ";
-        }
-        ss << "]";
+        // std::stringstream ss;
+        // ss << "[";
+        // for (auto token : token_strs) {
+        //     ss << "\"" << token << "\", ";
+        // }
+        // ss << "]";
        // LOG_DEBUG("split prompt \"%s\" to tokens %s", original_text.c_str(), ss.str().c_str());
        // printf("split prompt \"%s\" to tokens %s \n", original_text.c_str(), ss.str().c_str());
        return bpe_tokens;
@@ -194,10 +194,12 @@ public:
        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);

        x          = proj->forward(ctx, x);  // [ne3, ne2, ne1, dim_out*2]
-        auto x_vec = ggml_ext_chunk(ctx->ggml_ctx, x, 2, 0);
+        auto x_vec = ggml_ext_chunk(ctx->ggml_ctx, x, 2, 0, false);
        x          = x_vec[0];  // [ne3, ne2, ne1, dim_out]
        auto gate  = x_vec[1];  // [ne3, ne2, ne1, dim_out]

+        gate = ggml_cont(ctx->ggml_ctx, gate);
+
        gate = ggml_gelu_inplace(ctx->ggml_ctx, gate);

        x = ggml_mul(ctx->ggml_ctx, x, gate);  // [ne3, ne2, ne1, dim_out]
@@ -56,20 +56,26 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
    std::shared_ptr<CLIPTextModelRunner> text_model2;

    std::string trigger_word = "img";  // should be user settable
-    std::string embd_dir;
+    std::map<std::string, std::string> embedding_map;
    int32_t num_custom_embeddings   = 0;
    int32_t num_custom_embeddings_2 = 0;
    std::vector<uint8_t> token_embed_custom;
-    std::vector<std::string> readed_embeddings;
+    std::map<std::string, std::pair<int, int>> embedding_pos_map;

    FrozenCLIPEmbedderWithCustomWords(ggml_backend_t backend,
                                      bool offload_params_to_cpu,
                                      const String2TensorStorage& tensor_storage_map,
-                                      const std::string& embd_dir,
+                                      const std::map<std::string, std::string>& orig_embedding_map,
                                      SDVersion version = VERSION_SD1,
                                      PMVersion pv      = PM_VERSION_1)
-        : version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407), embd_dir(embd_dir) {
-        bool force_clip_f32 = embd_dir.size() > 0;
+        : version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407) {
+        for (const auto& kv : orig_embedding_map) {
+            std::string name = kv.first;
+            std::transform(name.begin(), name.end(), name.begin(), [](unsigned char c) { return std::tolower(c); });
+            embedding_map[name] = kv.second;
+            tokenizer.add_special_token(name);
+        }
+        bool force_clip_f32 = !embedding_map.empty();
        if (sd_version_is_sd1(version)) {
            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, true, force_clip_f32);
        } else if (sd_version_is_sd2(version)) {
@@ -117,14 +123,17 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
    }

    bool load_embedding(std::string embd_name, std::string embd_path, std::vector<int32_t>& bpe_tokens) {
-        // the order matters
        ModelLoader model_loader;
        if (!model_loader.init_from_file_and_convert_name(embd_path)) {
            LOG_ERROR("embedding '%s' failed", embd_name.c_str());
            return false;
        }
-        if (std::find(readed_embeddings.begin(), readed_embeddings.end(), embd_name) != readed_embeddings.end()) {
+        auto iter = embedding_pos_map.find(embd_name);
+        if (iter != embedding_pos_map.end()) {
            LOG_DEBUG("embedding already read in: %s", embd_name.c_str());
+            for (int i = iter->second.first; i < iter->second.second; i++) {
+                bpe_tokens.push_back(text_model->model.vocab_size + i);
+            }
            return true;
        }
        struct ggml_init_params params;
@@ -155,7 +164,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
            return true;
        };
        model_loader.load_tensors(on_load, 1);
-        readed_embeddings.push_back(embd_name);
+        int pos_start = num_custom_embeddings;
        if (embd) {
            int64_t hidden_size = text_model->model.hidden_size;
            token_embed_custom.resize(token_embed_custom.size() + ggml_nbytes(embd));
@@ -182,6 +191,11 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
            }
            LOG_DEBUG("embedding '%s' applied, custom embeddings: %i (text model 2)", embd_name.c_str(), num_custom_embeddings_2);
        }
+        int pos_end = num_custom_embeddings;
+        if (pos_end == pos_start) {
+            return false;
+        }
+        embedding_pos_map[embd_name] = std::pair{pos_start, pos_end};
        return true;
    }

@@ -196,25 +210,13 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {

    std::vector<int> convert_token_to_id(std::string text) {
        auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
-            size_t word_end       = str.find(",");
-            std::string embd_name = word_end == std::string::npos ? str : str.substr(0, word_end);
-            embd_name             = trim(embd_name);
-            std::string embd_path = get_full_path(embd_dir, embd_name + ".pt");
-            if (embd_path.size() == 0) {
-                embd_path = get_full_path(embd_dir, embd_name + ".ckpt");
+            auto iter = embedding_map.find(str);
+            if (iter == embedding_map.end()) {
+                return false;
            }
-            if (embd_path.size() == 0) {
-                embd_path = get_full_path(embd_dir, embd_name + ".safetensors");
-            }
-            if (embd_path.size() > 0) {
-                if (load_embedding(embd_name, embd_path, bpe_tokens)) {
-                    if (word_end != std::string::npos) {
-                        str = str.substr(word_end);
-                    } else {
-                        str = "";
-                    }
-                    return true;
-                }
+            std::string embedding_path = iter->second;
+            if (load_embedding(str, embedding_path, bpe_tokens)) {
+                return true;
            }
            return false;
        };
@@ -245,25 +247,13 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
        }

        auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
-            size_t word_end       = str.find(",");
-            std::string embd_name = word_end == std::string::npos ? str : str.substr(0, word_end);
-            embd_name             = trim(embd_name);
-            std::string embd_path = get_full_path(embd_dir, embd_name + ".pt");
-            if (embd_path.size() == 0) {
-                embd_path = get_full_path(embd_dir, embd_name + ".ckpt");
+            auto iter = embedding_map.find(str);
+            if (iter == embedding_map.end()) {
+                return false;
            }
-            if (embd_path.size() == 0) {
-                embd_path = get_full_path(embd_dir, embd_name + ".safetensors");
-            }
-            if (embd_path.size() > 0) {
-                if (load_embedding(embd_name, embd_path, bpe_tokens)) {
-                    if (word_end != std::string::npos) {
-                        str = str.substr(word_end);
-                    } else {
-                        str = "";
-                    }
-                    return true;
-                }
+            std::string embedding_path = iter->second;
+            if (load_embedding(str, embedding_path, bpe_tokens)) {
+                return true;
            }
            return false;
        };
@@ -376,25 +366,13 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
        }

        auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
-            size_t word_end       = str.find(",");
-            std::string embd_name = word_end == std::string::npos ? str : str.substr(0, word_end);
-            embd_name             = trim(embd_name);
-            std::string embd_path = get_full_path(embd_dir, embd_name + ".pt");
-            if (embd_path.size() == 0) {
-                embd_path = get_full_path(embd_dir, embd_name + ".ckpt");
+            auto iter = embedding_map.find(str);
+            if (iter == embedding_map.end()) {
+                return false;
            }
-            if (embd_path.size() == 0) {
-                embd_path = get_full_path(embd_dir, embd_name + ".safetensors");
-            }
-            if (embd_path.size() > 0) {
-                if (load_embedding(embd_name, embd_path, bpe_tokens)) {
-                    if (word_end != std::string::npos) {
-                        str = str.substr(word_end);
-                    } else {
-                        str = "";
-                    }
-                    return true;
-                }
+            std::string embedding_path = iter->second;
+            if (load_embedding(str, embedding_path, bpe_tokens)) {
+                return true;
            }
            return false;
        };
@@ -1638,7 +1616,7 @@ struct LLMEmbedder : public Conditioner {
        LLM::LLMArch arch = LLM::LLMArch::QWEN2_5_VL;
        if (sd_version_is_flux2(version)) {
            arch = LLM::LLMArch::MISTRAL_SMALL_3_2;
-        } else if (sd_version_is_z_image(version)) {
+        } else if (sd_version_is_z_image(version) || version == VERSION_OVIS_IMAGE) {
            arch = LLM::LLMArch::QWEN3;
        }
        if (arch == LLM::LLMArch::MISTRAL_SMALL_3_2) {
@@ -1728,6 +1706,7 @@ struct LLMEmbedder : public Conditioner {
        std::vector<std::pair<int, ggml_tensor*>> image_embeds;
        std::pair<int, int> prompt_attn_range;
        int prompt_template_encode_start_idx = 34;
+        int max_length                       = 0;
        std::set<int> out_layers;
        if (llm->enable_vision && conditioner_params.ref_images.size() > 0) {
            LOG_INFO("QwenImageEditPlusPipeline");
@@ -1825,6 +1804,17 @@ struct LLMEmbedder : public Conditioner {
            prompt_attn_range.second = prompt.size();

            prompt += "[/INST]";
+        } else if (version == VERSION_OVIS_IMAGE) {
+            prompt_template_encode_start_idx = 28;
+            max_length                       = prompt_template_encode_start_idx + 256;
+
+            prompt = "<|im_start|>user\nDescribe the image by detailing the color, quantity, text, shape, size, texture, spatial relationships of the objects and background:";
+
+            prompt_attn_range.first = static_cast<int>(prompt.size());
+            prompt += " " + conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+
+            prompt += "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n";
        } else {
            prompt_template_encode_start_idx = 34;

@@ -1837,7 +1827,7 @@ struct LLMEmbedder : public Conditioner {
            prompt += "<|im_end|>\n<|im_start|>assistant\n";
        }

-        auto tokens_and_weights = tokenize(prompt, prompt_attn_range, 0, false);
+        auto tokens_and_weights = tokenize(prompt, prompt_attn_range, max_length, max_length > 0);
        auto& tokens            = std::get<0>(tokens_and_weights);
        auto& weights           = std::get<1>(tokens_and_weights);

@@ -1870,9 +1860,13 @@ struct LLMEmbedder : public Conditioner {

        GGML_ASSERT(hidden_states->ne[1] > prompt_template_encode_start_idx);

-        int64_t zero_pad_len = 0;
+        int64_t min_length = 0;
        if (sd_version_is_flux2(version)) {
-            int64_t min_length = 512;
+            min_length = 512;
+        }
+
+        int64_t zero_pad_len = 0;
+        if (min_length > 0) {
            if (hidden_states->ne[1] - prompt_template_encode_start_idx < min_length) {
                zero_pad_len = min_length - hidden_states->ne[1] + prompt_template_encode_start_idx;
            }
@@ -1892,6 +1886,8 @@ struct LLMEmbedder : public Conditioner {
            ggml_ext_tensor_set_f32(new_hidden_states, value, i0, i1, i2, i3);
        });

+        // print_ggml_tensor(new_hidden_states);
+
        int64_t t1 = ggml_time_ms();
        LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
        return {new_hidden_states, nullptr, nullptr};
@@ -15,7 +15,7 @@ You can run Chroma using stable-diffusion.cpp with a GPU that has 6GB or even 4G
 You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF), this way you don't have to do the conversion yourself.

 ```
-.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
 ```

 ## Run
@@ -24,7 +24,7 @@ You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](h
 For example:

 ```
- .\bin\Release\sd.exe --diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask --clip-on-cpu
+ .\bin\Release\sd-cli.exe --diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask --clip-on-cpu
 ```

 ![](../assets/flux/chroma_v40.png)
@@ -12,7 +12,7 @@
 ## Examples

 ```
-.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Chroma1-Radiance-v0.4-Q8_0.gguf --t5xxl ..\..\ComfyUI\models\clip\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma  radiance cpp'" --cfg-scale 4.0 --sampling-method euler -v
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Chroma1-Radiance-v0.4-Q8_0.gguf --t5xxl ..\..\ComfyUI\models\clip\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma  radiance cpp'" --cfg-scale 4.0 --sampling-method euler -v
 ```

 <img alt="Chroma1-Radiance" src="../assets/flux/chroma1-radiance.png" />
@@ -9,7 +9,7 @@ docker build -t sd .
 ### Run

 ```shell
-docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
+docker run -v /path/to/models:/models -v /path/to/output/:/output sd-cli [args...]
 # For example
-# docker run -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
+# docker run -v ./models:/models -v ./build:/output sd-cli -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
 ```
@@ -5,5 +5,5 @@ You can use ESRGAN to upscale the generated images. At the moment, only the [Rea
 - Specify the model path using the `--upscale-model PATH` parameter. example:

 ```bash
-sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --upscale-model ../models/RealESRGAN_x4plus_anime_6B.pth
+sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --upscale-model ../models/RealESRGAN_x4plus_anime_6B.pth
 ```
@@ -17,7 +17,7 @@ You can download the preconverted gguf weights from [FLUX.1-dev-gguf](https://hu

 For example:
 ```
-.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
 ```

 ## Run
@@ -28,7 +28,7 @@ For example:
 For example:

 ```
- .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+ .\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
 ```

 Using formats of different precisions will yield results of varying quality.
@@ -44,7 +44,7 @@ Using formats of different precisions will yield results of varying quality.


 ```
- .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 --clip-on-cpu
+ .\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 --clip-on-cpu
 ```

 | q8_0  |
@@ -60,7 +60,7 @@ Since many flux LoRA training libraries have used various LoRA naming formats, i
 - LoRA model from https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main (using comfy converted version!!!)

 ```
-.\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models --clip-on-cpu
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models --clip-on-cpu
 ```

 ![output](../assets/flux/flux1-dev-q8_0%20with%20lora.png)
@@ -12,7 +12,7 @@
 ## Examples

 ```
-.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
 ```

 <img alt="flux2 example" src="../assets/flux2/example.png" />
@@ -82,4 +82,4 @@ cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_H
 cmake --build . --config Release
 ```

-If everything went OK, `build\bin\sd.exe` file should appear.
+If everything went OK, `build\bin\sd-cli.exe` file should appear.
@@ -16,7 +16,7 @@ You can run Kontext using stable-diffusion.cpp with a GPU that has 6GB or even 4
 You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF), this way you don't have to do the conversion yourself.

 ```
-.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
 ```

 ## Run
@@ -27,7 +27,7 @@ You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](ht
 For example:

 ```
- .\bin\Release\sd.exe -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+ .\bin\Release\sd-cli.exe -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
 ```


@@ -7,7 +7,7 @@
 Here's a simple example:

 ```
-./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:lcm-lora-sdv1-5:1>" --steps 4 --lora-model-dir ../models -v --cfg-scale 1
+./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:lcm-lora-sdv1-5:1>" --steps 4 --lora-model-dir ../models -v --cfg-scale 1
 ```

 | without LCM-LoRA (--cfg-scale 7)  | with LCM-LoRA (--cfg-scale 1)  |
@@ -7,7 +7,7 @@
 Here's a simple example:

 ```
-./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
+./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
 ```

 `../models/marblesh.safetensors` or `../models/marblesh.ckpt` will be applied to the model
@@ -0,0 +1,19 @@
+# How to Use
+
+## Download weights
+
+- Download Ovis-Image-7B
+    - safetensors: https://huggingface.co/Comfy-Org/Ovis-Image/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/leejet/Ovis-Image-7B-GGUF
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
+- Download Ovis 2.5
+    - safetensors: https://huggingface.co/Comfy-Org/Ovis-Image/tree/main/split_files/text_encoders
+
+## Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ovis_image-Q4_0.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\ovis_2.5.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="ovis image example" src="../assets/ovis_image/example.png" />
@@ -27,7 +27,7 @@ If on low memory GPUs (<= 8GB), recommend running with ```--vae-on-cpu``` option
 Example:

 ```bash
-bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --photo-maker ../models/photomaker-v1.safetensors --pm-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --pm-style-strength 10 --vae-on-cpu --steps 50
+bin/sd-cli -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --photo-maker ../models/photomaker-v1.safetensors --pm-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --pm-style-strength 10 --vae-on-cpu --steps 50
 ```

 ## PhotoMaker Version 2
@@ -23,5 +23,5 @@ You can also convert weights in the formats `ckpt/safetensors/diffusers` to gguf
 For example:

 ```sh
-./bin/sd -M convert -m ../models/v1-5-pruned-emaonly.safetensors -o  ../models/v1-5-pruned-emaonly.q8_0.gguf -v --type q8_0
+./bin/sd-cli -M convert -m ../models/v1-5-pruned-emaonly.safetensors -o  ../models/v1-5-pruned-emaonly.q8_0.gguf -v --type q8_0
 ```
@@ -14,7 +14,7 @@
 ## Examples

 ```
-.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
 ```

 <img alt="qwen example" src="../assets/qwen/example.png" />
@@ -20,7 +20,7 @@
 ### Qwen Image Edit

 ```
-.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
 ```

 <img alt="qwen_image_edit" src="../assets/qwen/qwen_image_edit.png" />
@@ -29,7 +29,7 @@
 ### Qwen Image Edit 2509

 ```
-.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --llm_vision ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --llm_vision ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
 ```

 <img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2509.png" />
@@ -9,12 +9,12 @@
 ### txt2img example

 ```sh
-./bin/sd -m ../models/sd-v1-4.ckpt -p "a lovely cat"
-# ./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
-# ./bin/sd -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
-# ./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
-# ./bin/sd --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
-# ./bin/sd -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+./bin/sd-cli -m ../models/sd-v1-4.ckpt -p "a lovely cat"
+# ./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
+# ./bin/sd-cli -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
+# ./bin/sd-cli -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd-cli --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd-cli -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
 ```

 Using formats of different precisions will yield results of varying quality.
@@ -29,7 +29,7 @@ Using formats of different precisions will yield results of varying quality.


 ```
-./bin/sd -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
+./bin/sd-cli -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
 ```

 <p align="center">
@@ -14,7 +14,7 @@
 For example:

 ```
-.\bin\Release\sd.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+.\bin\Release\sd-cli.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
 ```

 ![](../assets/sd3.5_large.png)
@@ -13,5 +13,5 @@ curl -L -O https://huggingface.co/madebyollin/taesd/resolve/main/diffusion_pytor
 - Specify the model path using the `--taesd PATH` parameter. example:

 ```bash
-sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --taesd ../models/diffusion_pytorch_model.safetensors
+sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --taesd ../models/diffusion_pytorch_model.safetensors
 ```
@@ -52,7 +52,7 @@
 ### Wan2.1 T2V 1.3B

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.1_1.3B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -60,7 +60,7 @@
 ### Wan2.1 T2V 14B

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.1_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -70,7 +70,7 @@
 ### Wan2.1 I2V 14B

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.1_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -78,7 +78,7 @@
 ### Wan2.2 T2V A14B

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -86,7 +86,7 @@
 ### Wan2.2 I2V A14B

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -94,7 +94,7 @@
 ### Wan2.2 T2V A14B T2I

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
 ```

 <img width="832" height="480" alt="Wan2 2_14B_t2i" src="../assets/wan/Wan2.2_14B_t2i.png" />
@@ -102,7 +102,7 @@
 ### Wan2.2 T2V 14B with Lora

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\..\ComfyUI\models\loras --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\..\ComfyUI\models\loras --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_14B_t2v_lora.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -114,7 +114,7 @@
 #### T2V

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_5B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -122,7 +122,7 @@
 #### I2V

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_5B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -130,7 +130,7 @@
 ### Wan2.1 FLF2V 14B

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
 ```


@@ -139,7 +139,7 @@
 ### Wan2.2 FLF2V 14B

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
 ```

 <video src=../assets/wan/Wan2.2_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -149,7 +149,7 @@
 #### T2V

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_1.3B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -158,7 +158,7 @@
 #### R2V

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_1.3B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -169,7 +169,7 @@
 ```
 mkdir post+depth
 ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\frame_%04d.jpg
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_1.3B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -179,7 +179,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### T2V

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_14B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -188,7 +188,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### R2V

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_14B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -198,7 +198,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### V2V

 ```
-.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
 ```

 <video src=../assets/wan/Wan2.1_14B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -16,7 +16,7 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 ## Examples

 ```
-.\bin\Release\sd.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
+.\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
 ```

 <img width="256" alt="z-image example" src="../assets/z_image/q3_K.png" />
@@ -156,9 +156,10 @@ struct ESRGAN : public GGMLRunner {

    ESRGAN(ggml_backend_t backend,
           bool offload_params_to_cpu,
+           int tile_size                                  = 128,
           const String2TensorStorage& tensor_storage_map = {})
        : GGMLRunner(backend, offload_params_to_cpu) {
-        // rrdb_net will be created in load_from_file
+        this->tile_size = tile_size;
    }

    std::string get_desc() override {
@@ -1,3 +1,4 @@
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})

-add_subdirectory(cli)
+add_subdirectory(cli)
+add_subdirectory(server)
@@ -1,4 +1,4 @@
-set(TARGET sd)
+set(TARGET sd-cli)

 add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
@@ -1,7 +1,7 @@
 # Run

 ```
-usage: ./bin/sd  [options]
+usage: ./bin/sd-cli  [options]

 CLI Options:
  -o, --output <string>       path to write result image to (default: ./output.png)
@@ -121,6 +121,7 @@ Generation Options:
                                           ddim_trailing, tcd] default: euler for Flux/SD3/Wan, euler_a otherwise
  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, lcm],
                                           default: discrete
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
  --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
  --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
  -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
@@ -0,0 +1,6 @@
+set(TARGET sd-server)
+
+add_executable(${TARGET} main.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)
@@ -0,0 +1,123 @@
+# Run
+
+```
+usage: ./bin/sd-server  [options]
+
+Svr Options:
+  -l, --listen-ip <string>    server listen ip (default: 127.0.0.1)
+  --listen-port <int>         server listen port (default: 1234)
+  -v, --verbose               print extra info
+  --color                     colors the logging tags according to level
+  -h, --help                  show this help message and exit
+
+Context Options:
+  -m, --model <string>                     path to full model
+  --clip_l <string>                        path to the clip-l text encoder
+  --clip_g <string>                        path to the clip-g text encoder
+  --clip_vision <string>                   path to the clip-vision encoder
+  --t5xxl <string>                         path to the t5xxl text encoder
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
+  --llm_vision <string>                    path to the llm vit
+  --qwen2vl <string>                       alias of --llm. Deprecated.
+  --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
+  --diffusion-model <string>               path to the standalone diffusion model
+  --high-noise-diffusion-model <string>    path to the standalone high noise diffusion model
+  --vae <string>                           path to standalone vae model
+  --taesd <string>                         path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
+  --control-net <string>                   path to control net model
+  --embd-dir <string>                      embeddings directory
+  --lora-model-dir <string>                lora model directory
+  --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
+  --photo-maker <string>                   path to PHOTOMAKER model
+  --upscale-model <string>                 path to esrgan model.
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
+                                           CPU physical cores
+  --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
+  --vae-tiling                             process vae in tiles to reduce memory usage
+  --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
+  --control-net-cpu                        keep controlnet in cpu (for low vram)
+  --clip-on-cpu                            keep clip in cpu (for low vram)
+  --vae-on-cpu                             keep vae in cpu (for low vram)
+  --diffusion-fa                           use flash attention in the diffusion model
+  --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
+  --vae-conv-direct                        use ggml_conv2d_direct in the vae model
+  --chroma-disable-dit-mask                disable dit mask for chroma
+  --chroma-enable-t5-mask                  enable t5 mask for chroma
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
+                                           type of the weight file
+  --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
+  --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
+                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
+                                           immediately will be used.The immediately mode may have precision and
+                                           compatibility issues with quantized parameters, but it usually offers faster inference
+                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
+                                           other hand, is exactly the opposite.
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
+                                           (overrides --vae-tile-size)
+
+Default Generation Options:
+  -p, --prompt <string>                    the prompt to render
+  -n, --negative-prompt <string>           the negative prompt (default: "")
+  -i, --init-img <string>                  path to the init image
+  --end-img <string>                       path to the end image, required by flf2v
+  --mask <string>                          path to the mask image
+  --control-image <string>                 path to control image, control net
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
+                                           lexicographical (character) order. For example, if the control video path is
+                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
+  --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
+  --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
+  -H, --height <int>                       image height, in pixel space (default: 512)
+  -W, --width <int>                        image width, in pixel space (default: 512)
+  --steps <int>                            number of sample steps (default: 20)
+  --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
+                                           will be 1 for SD1.x, 2 for SD2.x
+  -b, --batch-count <int>                  batch count
+  --video-frames <int>                     video frames (default: 1)
+  --fps <int>                              fps (default: 24)
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
+                                           NitroSD-Vibrant
+  --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
+  --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
+  --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
+                                           medium
+  --skip-layer-start <float>               SLG enabling point (default: 0.01)
+  --skip-layer-end <float>                 SLG disabling point (default: 0.2)
+  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
+  --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
+  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+  --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
+  --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
+  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM and TCD (default: 0)
+  --strength <float>                       strength for noising/unnoising (default: 0.75)
+  --pm-style-strength <float>
+  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
+  --vace-strength <float>                  wan vace strength
+  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
+  --disable-auto-resize-ref-image          disable auto resize of ref images
+  -s, --seed                               RNG seed (default: 42, use random seed for < 0)
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
+                                           tcd] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
+                                           ddim_trailing, tcd] default: euler for Flux/SD3/Wan, euler_a otherwise
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, lcm],
+                                           default: discrete
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
+  --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
+  --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
+  -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
+  --easycache                              enable EasyCache for DiT models with optional "threshold,start_percent,end_percent" (default: 0.2,0.15,0.95)
+```
@@ -0,0 +1,744 @@
+// main.cpp
+#include <chrono>
+#include <filesystem>
+#include <fstream>
+#include <iomanip>
+#include <iostream>
+#include <mutex>
+#include <sstream>
+#include <vector>
+
+#include "httplib.h"
+#include "stable-diffusion.h"
+
+#include "common/common.hpp"
+
+namespace fs = std::filesystem;
+
+// ----------------------- helpers -----------------------
+static const std::string base64_chars =
+    "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+    "abcdefghijklmnopqrstuvwxyz"
+    "0123456789+/";
+
+std::string base64_encode(const std::vector<uint8_t>& bytes) {
+    std::string ret;
+    int val = 0, valb = -6;
+    for (uint8_t c : bytes) {
+        val = (val << 8) + c;
+        valb += 8;
+        while (valb >= 0) {
+            ret.push_back(base64_chars[(val >> valb) & 0x3F]);
+            valb -= 6;
+        }
+    }
+    if (valb > -6)
+        ret.push_back(base64_chars[((val << 8) >> (valb + 8)) & 0x3F]);
+    while (ret.size() % 4)
+        ret.push_back('=');
+    return ret;
+}
+
+inline bool is_base64(unsigned char c) {
+    return (isalnum(c) || (c == '+') || (c == '/'));
+}
+
+std::vector<uint8_t> base64_decode(const std::string& encoded_string) {
+    int in_len = encoded_string.size();
+    int i      = 0;
+    int j      = 0;
+    int in_    = 0;
+    uint8_t char_array_4[4], char_array_3[3];
+    std::vector<uint8_t> ret;
+
+    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) {
+        char_array_4[i++] = encoded_string[in_];
+        in_++;
+        if (i == 4) {
+            for (i = 0; i < 4; i++)
+                char_array_4[i] = static_cast<uint8_t>(base64_chars.find(char_array_4[i]));
+
+            char_array_3[0] = (char_array_4[0] << 2) + ((char_array_4[1] & 0x30) >> 4);
+            char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+            char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3];
+
+            for (i = 0; i < 3; i++)
+                ret.push_back(char_array_3[i]);
+            i = 0;
+        }
+    }
+
+    if (i) {
+        for (j = i; j < 4; j++)
+            char_array_4[j] = 0;
+
+        for (j = 0; j < 4; j++)
+            char_array_4[j] = static_cast<uint8_t>(base64_chars.find(char_array_4[j]));
+
+        char_array_3[0] = (char_array_4[0] << 2) + ((char_array_4[1] & 0x30) >> 4);
+        char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+        char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3];
+
+        for (j = 0; j < i - 1; j++)
+            ret.push_back(char_array_3[j]);
+    }
+
+    return ret;
+}
+
+std::string iso_timestamp_now() {
+    using namespace std::chrono;
+    auto now      = system_clock::now();
+    std::time_t t = system_clock::to_time_t(now);
+    std::tm tm{};
+#ifdef _MSC_VER
+    gmtime_s(&tm, &t);
+#else
+    gmtime_r(&t, &tm);
+#endif
+    std::ostringstream oss;
+    oss << std::put_time(&tm, "%Y-%m-%dT%H:%M:%SZ");
+    return oss.str();
+}
+
+struct SDSvrParams {
+    std::string listen_ip = "127.0.0.1";
+    int listen_port       = 1234;
+    bool normal_exit      = false;
+    bool verbose          = false;
+    bool color            = false;
+
+    ArgOptions get_options() {
+        ArgOptions options;
+
+        options.string_options = {
+            {"-l",
+             "--listen-ip",
+             "server listen ip (default: 127.0.0.1)",
+             &listen_ip}};
+
+        options.int_options = {
+            {"",
+             "--listen-port",
+             "server listen port (default: 1234)",
+             &listen_port},
+        };
+
+        options.bool_options = {
+            {"-v",
+             "--verbose",
+             "print extra info",
+             true, &verbose},
+            {"",
+             "--color",
+             "colors the logging tags according to level",
+             true, &color},
+        };
+
+        auto on_help_arg = [&](int argc, const char** argv, int index) {
+            normal_exit = true;
+            return -1;
+        };
+
+        options.manual_options = {
+            {"-h",
+             "--help",
+             "show this help message and exit",
+             on_help_arg},
+        };
+        return options;
+    };
+
+    bool process_and_check() {
+        if (listen_ip.empty()) {
+            fprintf(stderr, "error: the following arguments are required: listen_ip\n");
+            return false;
+        }
+
+        if (listen_port < 0 || listen_port > 65535) {
+            fprintf(stderr, "error: listen_port should be in the range [0, 65535]\n");
+            return false;
+        }
+        return true;
+    }
+
+    std::string to_string() const {
+        std::ostringstream oss;
+        oss << "SDSvrParams {\n"
+            << "  listen_ip: " << listen_ip << ",\n"
+            << "  listen_port: \"" << listen_port << "\",\n"
+            << "}";
+        return oss.str();
+    }
+};
+
+void print_usage(int argc, const char* argv[], const std::vector<ArgOptions>& options_list) {
+    std::cout << version_string() << "\n";
+    std::cout << "Usage: " << argv[0] << " [options]\n\n";
+    std::cout << "Svr Options:\n";
+    options_list[0].print();
+    std::cout << "\nContext Options:\n";
+    options_list[1].print();
+    std::cout << "\nDefault Generation Options:\n";
+    options_list[2].print();
+}
+
+void parse_args(int argc, const char** argv, SDSvrParams& svr_params, SDContextParams& ctx_params, SDGenerationParams& default_gen_params) {
+    std::vector<ArgOptions> options_vec = {svr_params.get_options(), ctx_params.get_options(), default_gen_params.get_options()};
+
+    if (!parse_options(argc, argv, options_vec)) {
+        print_usage(argc, argv, options_vec);
+        exit(svr_params.normal_exit ? 0 : 1);
+    }
+
+    if (!svr_params.process_and_check() ||
+        !ctx_params.process_and_check(IMG_GEN) ||
+        !default_gen_params.process_and_check(IMG_GEN, ctx_params.lora_model_dir)) {
+        print_usage(argc, argv, options_vec);
+        exit(1);
+    }
+}
+
+std::string extract_and_remove_sd_cpp_extra_args(std::string& text) {
+    std::regex re("<sd_cpp_extra_args>(.*?)</sd_cpp_extra_args>");
+    std::smatch match;
+
+    std::string extracted;
+    if (std::regex_search(text, match, re)) {
+        extracted = match[1].str();
+        text      = std::regex_replace(text, re, "");
+    }
+    return extracted;
+}
+
+enum class ImageFormat { JPEG,
+                         PNG };
+
+std::vector<uint8_t> write_image_to_vector(
+    ImageFormat format,
+    const uint8_t* image,
+    int width,
+    int height,
+    int channels,
+    int quality = 90) {
+    std::vector<uint8_t> buffer;
+
+    auto write_func = [&buffer](void* context, void* data, int size) {
+        uint8_t* src = reinterpret_cast<uint8_t*>(data);
+        buffer.insert(buffer.end(), src, src + size);
+    };
+
+    struct ContextWrapper {
+        decltype(write_func)& func;
+    } ctx{write_func};
+
+    auto c_func = [](void* context, void* data, int size) {
+        auto* wrapper = reinterpret_cast<ContextWrapper*>(context);
+        wrapper->func(context, data, size);
+    };
+
+    int result = 0;
+    switch (format) {
+        case ImageFormat::JPEG:
+            result = stbi_write_jpg_to_func(c_func, &ctx, width, height, channels, image, quality);
+            break;
+        case ImageFormat::PNG:
+            result = stbi_write_png_to_func(c_func, &ctx, width, height, channels, image, width * channels);
+            break;
+        default:
+            throw std::runtime_error("invalid image format");
+    }
+
+    if (!result) {
+        throw std::runtime_error("write imgage to mem failed");
+    }
+
+    return buffer;
+}
+
+/* Enables Printing the log level tag in color using ANSI escape codes */
+void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
+    SDSvrParams* svr_params = (SDSvrParams*)data;
+    int tag_color;
+    const char* level_str;
+    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
+
+    if (!log || (!svr_params->verbose && level <= SD_LOG_DEBUG)) {
+        return;
+    }
+
+    switch (level) {
+        case SD_LOG_DEBUG:
+            tag_color = 37;
+            level_str = "DEBUG";
+            break;
+        case SD_LOG_INFO:
+            tag_color = 34;
+            level_str = "INFO";
+            break;
+        case SD_LOG_WARN:
+            tag_color = 35;
+            level_str = "WARN";
+            break;
+        case SD_LOG_ERROR:
+            tag_color = 31;
+            level_str = "ERROR";
+            break;
+        default: /* Potential future-proofing */
+            tag_color = 33;
+            level_str = "?????";
+            break;
+    }
+
+    if (svr_params->color == true) {
+        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
+    } else {
+        fprintf(out_stream, "[%-5s] ", level_str);
+    }
+    fputs(log, out_stream);
+    fflush(out_stream);
+}
+
+int main(int argc, const char** argv) {
+    SDSvrParams svr_params;
+    SDContextParams ctx_params;
+    SDGenerationParams default_gen_params;
+    parse_args(argc, argv, svr_params, ctx_params, default_gen_params);
+
+    sd_set_log_callback(sd_log_cb, (void*)&svr_params);
+
+    if (svr_params.verbose) {
+        printf("%s", sd_get_system_info());
+        printf("%s\n", svr_params.to_string().c_str());
+        printf("%s\n", ctx_params.to_string().c_str());
+        printf("%s\n", default_gen_params.to_string().c_str());
+    }
+
+    sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(false, false, false);
+    sd_ctx_t* sd_ctx              = new_sd_ctx(&sd_ctx_params);
+
+    if (sd_ctx == nullptr) {
+        printf("new_sd_ctx_t failed\n");
+        return 1;
+    }
+
+    std::mutex sd_ctx_mutex;
+
+    httplib::Server svr;
+
+    svr.set_pre_routing_handler([](const httplib::Request& req, httplib::Response& res) {
+        std::string origin = req.get_header_value("Origin");
+        if (origin.empty()) {
+            origin = "*";
+        }
+        res.set_header("Access-Control-Allow-Origin", origin);
+        res.set_header("Access-Control-Allow-Credentials", "true");
+        res.set_header("Access-Control-Allow-Methods", "*");
+        res.set_header("Access-Control-Allow-Headers", "*");
+
+        if (req.method == "OPTIONS") {
+            res.status = 204;
+            return httplib::Server::HandlerResponse::Handled;
+        }
+        return httplib::Server::HandlerResponse::Unhandled;
+    });
+
+    // health
+    svr.Get("/", [&](const httplib::Request&, httplib::Response& res) {
+        res.set_content(R"({"ok":true,"service":"sd-cpp-http"})", "application/json");
+    });
+
+    // models endpoint (minimal)
+    svr.Get("/v1/models", [&](const httplib::Request&, httplib::Response& res) {
+        json r;
+        r["data"] = json::array();
+        r["data"].push_back({{"id", "sd-cpp-local"}, {"object", "model"}, {"owned_by", "local"}});
+        res.set_content(r.dump(), "application/json");
+    });
+
+    // core endpoint: /v1/images/generations
+    svr.Post("/v1/images/generations", [&](const httplib::Request& req, httplib::Response& res) {
+        try {
+            if (req.body.empty()) {
+                res.status = 400;
+                res.set_content(R"({"error":"empty body"})", "application/json");
+                return;
+            }
+
+            json j                    = json::parse(req.body);
+            std::string prompt        = j.value("prompt", "");
+            int n                     = std::max(1, j.value("n", 1));
+            std::string size          = j.value("size", "");
+            std::string output_format = j.value("output_format", "png");
+            int output_compression    = j.value("output_compression", 100);
+            int width                 = 512;
+            int height                = 512;
+            if (!size.empty()) {
+                auto pos = size.find('x');
+                if (pos != std::string::npos) {
+                    try {
+                        width  = std::stoi(size.substr(0, pos));
+                        height = std::stoi(size.substr(pos + 1));
+                    } catch (...) {
+                    }
+                }
+            }
+
+            if (prompt.empty()) {
+                res.status = 400;
+                res.set_content(R"({"error":"prompt required"})", "application/json");
+                return;
+            }
+
+            std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(prompt);
+
+            if (output_format != "png" && output_format != "jpeg") {
+                res.status = 400;
+                res.set_content(R"({"error":"invalid output_format, must be one of [png, jpeg]"})", "application/json");
+                return;
+            }
+            if (n <= 0)
+                n = 1;
+            if (n > 8)
+                n = 8;  // safety
+            if (output_compression > 100) {
+                output_compression = 100;
+            }
+            if (output_compression < 0) {
+                output_compression = 0;
+            }
+
+            json out;
+            out["created"]       = iso_timestamp_now();
+            out["data"]          = json::array();
+            out["output_format"] = output_format;
+
+            SDGenerationParams gen_params = default_gen_params;
+            gen_params.prompt             = prompt;
+            gen_params.width              = width;
+            gen_params.height             = height;
+            gen_params.batch_count        = n;
+
+            if (!sd_cpp_extra_args_str.empty() && !gen_params.from_json_str(sd_cpp_extra_args_str)) {
+                res.status = 400;
+                res.set_content(R"({"error":"invalid sd_cpp_extra_args"})", "application/json");
+                return;
+            }
+
+            if (!gen_params.process_and_check(IMG_GEN, "")) {
+                res.status = 400;
+                res.set_content(R"({"error":"invalid params"})", "application/json");
+                return;
+            }
+
+            if (svr_params.verbose) {
+                printf("%s\n", gen_params.to_string().c_str());
+            }
+
+            sd_image_t init_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+            sd_image_t control_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+            sd_image_t mask_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 1, nullptr};
+            std::vector<sd_image_t> pmid_images;
+
+            sd_img_gen_params_t img_gen_params = {
+                gen_params.lora_vec.data(),
+                static_cast<uint32_t>(gen_params.lora_vec.size()),
+                gen_params.prompt.c_str(),
+                gen_params.negative_prompt.c_str(),
+                gen_params.clip_skip,
+                init_image,
+                nullptr,
+                0,
+                gen_params.auto_resize_ref_image,
+                gen_params.increase_ref_index,
+                mask_image,
+                gen_params.width,
+                gen_params.height,
+                gen_params.sample_params,
+                gen_params.strength,
+                gen_params.seed,
+                gen_params.batch_count,
+                control_image,
+                gen_params.control_strength,
+                {
+                    pmid_images.data(),
+                    (int)pmid_images.size(),
+                    gen_params.pm_id_embed_path.c_str(),
+                    gen_params.pm_style_strength,
+                },  // pm_params
+                ctx_params.vae_tiling_params,
+                gen_params.easycache_params,
+            };
+
+            sd_image_t* results = nullptr;
+            int num_results     = 0;
+
+            {
+                std::lock_guard<std::mutex> lock(sd_ctx_mutex);
+                results     = generate_image(sd_ctx, &img_gen_params);
+                num_results = gen_params.batch_count;
+            }
+
+            for (int i = 0; i < num_results; i++) {
+                if (results[i].data == nullptr) {
+                    continue;
+                }
+                auto image_bytes = write_image_to_vector(output_format == "jpeg" ? ImageFormat::JPEG : ImageFormat::PNG,
+                                                         results[i].data,
+                                                         results[i].width,
+                                                         results[i].height,
+                                                         results[i].channel,
+                                                         output_compression);
+                if (image_bytes.empty()) {
+                    printf("write image to mem failed\n");
+                    continue;
+                }
+
+                // base64 encode
+                std::string b64 = base64_encode(image_bytes);
+                json item;
+                item["b64_json"] = b64;
+                out["data"].push_back(item);
+            }
+
+            res.set_content(out.dump(), "application/json");
+            res.status = 200;
+
+        } catch (const std::exception& e) {
+            res.status = 500;
+            json err;
+            err["error"]   = "server_error";
+            err["message"] = e.what();
+            res.set_content(err.dump(), "application/json");
+        }
+    });
+
+    svr.Post("/v1/images/edits", [&](const httplib::Request& req, httplib::Response& res) {
+        try {
+            if (!req.is_multipart_form_data()) {
+                res.status = 400;
+                res.set_content(R"({"error":"Content-Type must be multipart/form-data"})", "application/json");
+                return;
+            }
+
+            std::string prompt = req.form.get_field("prompt");
+            if (prompt.empty()) {
+                res.status = 400;
+                res.set_content(R"({"error":"prompt required"})", "application/json");
+                return;
+            }
+
+            std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(prompt);
+
+            size_t image_count = req.form.get_file_count("image[]");
+            if (image_count == 0) {
+                res.status = 400;
+                res.set_content(R"({"error":"at least one image[] required"})", "application/json");
+                return;
+            }
+
+            std::vector<std::vector<uint8_t>> images_bytes;
+            for (size_t i = 0; i < image_count; i++) {
+                auto file = req.form.get_file("image[]", i);
+                images_bytes.emplace_back(file.content.begin(), file.content.end());
+            }
+
+            std::vector<uint8_t> mask_bytes;
+            if (req.form.has_field("mask")) {
+                auto file = req.form.get_file("mask");
+                mask_bytes.assign(file.content.begin(), file.content.end());
+            }
+
+            int n = 1;
+            if (req.form.has_field("n")) {
+                try {
+                    n = std::stoi(req.form.get_field("n"));
+                } catch (...) {
+                }
+            }
+            n = std::clamp(n, 1, 8);
+
+            std::string size = req.form.get_field("size");
+            int width = 512, height = 512;
+            if (!size.empty()) {
+                auto pos = size.find('x');
+                if (pos != std::string::npos) {
+                    try {
+                        width  = std::stoi(size.substr(0, pos));
+                        height = std::stoi(size.substr(pos + 1));
+                    } catch (...) {
+                    }
+                }
+            }
+
+            std::string output_format = "png";
+            if (req.form.has_field("output_format"))
+                output_format = req.form.get_field("output_format");
+            if (output_format != "png" && output_format != "jpeg") {
+                res.status = 400;
+                res.set_content(R"({"error":"invalid output_format, must be one of [png, jpeg]"})", "application/json");
+                return;
+            }
+
+            std::string output_compression_str = req.form.get_field("output_compression");
+            int output_compression             = 100;
+            try {
+                output_compression = std::stoi(output_compression_str);
+            } catch (...) {
+            }
+            if (output_compression > 100) {
+                output_compression = 100;
+            }
+            if (output_compression < 0) {
+                output_compression = 0;
+            }
+
+            SDGenerationParams gen_params = default_gen_params;
+            gen_params.prompt             = prompt;
+            gen_params.width              = width;
+            gen_params.height             = height;
+            gen_params.batch_count        = n;
+
+            if (!sd_cpp_extra_args_str.empty() && !gen_params.from_json_str(sd_cpp_extra_args_str)) {
+                res.status = 400;
+                res.set_content(R"({"error":"invalid sd_cpp_extra_args"})", "application/json");
+                return;
+            }
+
+            if (!gen_params.process_and_check(IMG_GEN, "")) {
+                res.status = 400;
+                res.set_content(R"({"error":"invalid params"})", "application/json");
+                return;
+            }
+
+            if (svr_params.verbose) {
+                printf("%s\n", gen_params.to_string().c_str());
+            }
+
+            sd_image_t init_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+            sd_image_t control_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+            std::vector<sd_image_t> pmid_images;
+
+            std::vector<sd_image_t> ref_images;
+            ref_images.reserve(images_bytes.size());
+            for (auto& bytes : images_bytes) {
+                int img_w           = width;
+                int img_h           = height;
+                uint8_t* raw_pixels = load_image_from_memory(
+                    reinterpret_cast<const char*>(bytes.data()),
+                    bytes.size(),
+                    img_w, img_h,
+                    width, height, 3);
+
+                if (!raw_pixels) {
+                    continue;
+                }
+
+                sd_image_t img{(uint32_t)img_w, (uint32_t)img_h, 3, raw_pixels};
+                ref_images.push_back(img);
+            }
+
+            sd_image_t mask_image = {0};
+            if (!mask_bytes.empty()) {
+                int mask_w        = width;
+                int mask_h        = height;
+                uint8_t* mask_raw = load_image_from_memory(
+                    reinterpret_cast<const char*>(mask_bytes.data()),
+                    mask_bytes.size(),
+                    mask_w, mask_h,
+                    width, height, 1);
+                mask_image = {(uint32_t)mask_w, (uint32_t)mask_h, 1, mask_raw};
+            } else {
+                mask_image.width   = width;
+                mask_image.height  = height;
+                mask_image.channel = 1;
+                mask_image.data    = nullptr;
+            }
+
+            sd_img_gen_params_t img_gen_params = {
+                gen_params.lora_vec.data(),
+                static_cast<uint32_t>(gen_params.lora_vec.size()),
+                gen_params.prompt.c_str(),
+                gen_params.negative_prompt.c_str(),
+                gen_params.clip_skip,
+                init_image,
+                ref_images.data(),
+                (int)ref_images.size(),
+                gen_params.auto_resize_ref_image,
+                gen_params.increase_ref_index,
+                mask_image,
+                gen_params.width,
+                gen_params.height,
+                gen_params.sample_params,
+                gen_params.strength,
+                gen_params.seed,
+                gen_params.batch_count,
+                control_image,
+                gen_params.control_strength,
+                {
+                    pmid_images.data(),
+                    (int)pmid_images.size(),
+                    gen_params.pm_id_embed_path.c_str(),
+                    gen_params.pm_style_strength,
+                },  // pm_params
+                ctx_params.vae_tiling_params,
+                gen_params.easycache_params,
+            };
+
+            sd_image_t* results = nullptr;
+            int num_results     = 0;
+
+            {
+                std::lock_guard<std::mutex> lock(sd_ctx_mutex);
+                results     = generate_image(sd_ctx, &img_gen_params);
+                num_results = gen_params.batch_count;
+            }
+
+            json out;
+            out["created"]       = iso_timestamp_now();
+            out["data"]          = json::array();
+            out["output_format"] = output_format;
+
+            for (int i = 0; i < num_results; i++) {
+                if (results[i].data == nullptr)
+                    continue;
+                auto image_bytes = write_image_to_vector(output_format == "jpeg" ? ImageFormat::JPEG : ImageFormat::PNG,
+                                                         results[i].data,
+                                                         results[i].width,
+                                                         results[i].height,
+                                                         results[i].channel,
+                                                         output_compression);
+                std::string b64 = base64_encode(image_bytes);
+                json item;
+                item["b64_json"] = b64;
+                out["data"].push_back(item);
+            }
+
+            res.set_content(out.dump(), "application/json");
+            res.status = 200;
+
+            if (init_image.data) {
+                stbi_image_free(init_image.data);
+            }
+            if (mask_image.data) {
+                stbi_image_free(mask_image.data);
+            }
+            for (auto ref_image : ref_images) {
+                stbi_image_free(ref_image.data);
+            }
+        } catch (const std::exception& e) {
+            res.status = 500;
+            json err;
+            err["error"]   = "server_error";
+            err["message"] = e.what();
+            res.set_content(err.dump(), "application/json");
+        }
+    });
+
+    printf("listening on: %s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
+    svr.listen(svr_params.listen_ip, svr_params.listen_port);
+
+    // cleanup
+    free_sd_ctx(sd_ctx);
+    return 0;
+}
@@ -134,6 +134,54 @@ namespace Flux {
        }
    };

+    struct MLP : public UnaryBlock {
+        bool use_mlp_silu_act;
+
+    public:
+        MLP(int64_t hidden_size, int64_t intermediate_size, bool use_mlp_silu_act = false, bool bias = false)
+            : use_mlp_silu_act(use_mlp_silu_act) {
+            int64_t mlp_mult_factor = use_mlp_silu_act ? 2 : 1;
+            blocks["0"]             = std::make_shared<Linear>(hidden_size, intermediate_size * mlp_mult_factor, bias);
+            blocks["2"]             = std::make_shared<Linear>(intermediate_size, hidden_size, bias);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["0"]);
+            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
+
+            x = mlp_0->forward(ctx, x);
+            if (use_mlp_silu_act) {
+                x = ggml_ext_silu_act(ctx->ggml_ctx, x);
+            } else {
+                x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+            }
+            x = mlp_2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct YakMLP : public UnaryBlock {
+    public:
+        YakMLP(int64_t hidden_size, int64_t intermediate_size, bool bias = true) {
+            blocks["gate_proj"] = std::make_shared<Linear>(hidden_size, intermediate_size, bias);
+            blocks["up_proj"]   = std::make_shared<Linear>(hidden_size, intermediate_size, bias);
+            blocks["down_proj"] = std::make_shared<Linear>(intermediate_size, hidden_size, bias);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+            auto gate_proj = std::dynamic_pointer_cast<Linear>(blocks["gate_proj"]);
+            auto up_proj   = std::dynamic_pointer_cast<Linear>(blocks["up_proj"]);
+            auto down_proj = std::dynamic_pointer_cast<Linear>(blocks["down_proj"]);
+
+            auto gate = gate_proj->forward(ctx, x);
+            gate      = ggml_silu_inplace(ctx->ggml_ctx, gate);
+            x         = up_proj->forward(ctx, x);
+            x         = ggml_mul(ctx->ggml_ctx, x, gate);
+            x         = down_proj->forward(ctx, x);
+            return x;
+        }
+    };
+
    struct ModulationOut {
        ggml_tensor* shift = nullptr;
        ggml_tensor* scale = nullptr;
@@ -199,7 +247,6 @@ namespace Flux {
    struct DoubleStreamBlock : public GGMLBlock {
        bool prune_mod;
        int idx = 0;
-        bool use_mlp_silu_act;

    public:
        DoubleStreamBlock(int64_t hidden_size,
@@ -210,10 +257,10 @@ namespace Flux {
                          bool prune_mod        = false,
                          bool share_modulation = false,
                          bool mlp_proj_bias    = true,
+                          bool use_yak_mlp      = false,
                          bool use_mlp_silu_act = false)
-            : idx(idx), prune_mod(prune_mod), use_mlp_silu_act(use_mlp_silu_act) {
-            int64_t mlp_hidden_dim  = hidden_size * mlp_ratio;
-            int64_t mlp_mult_factor = use_mlp_silu_act ? 2 : 1;
+            : idx(idx), prune_mod(prune_mod) {
+            int64_t mlp_hidden_dim = hidden_size * mlp_ratio;

            if (!prune_mod && !share_modulation) {
                blocks["img_mod"] = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
@@ -222,9 +269,11 @@ namespace Flux {
            blocks["img_attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias, mlp_proj_bias));

            blocks["img_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
-            blocks["img_mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, mlp_hidden_dim * mlp_mult_factor, mlp_proj_bias));
-            // img_mlp.1 is nn.GELU(approximate="tanh")
-            blocks["img_mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(mlp_hidden_dim, hidden_size, mlp_proj_bias));
+            if (use_yak_mlp) {
+                blocks["img_mlp"] = std::shared_ptr<GGMLBlock>(new YakMLP(hidden_size, mlp_hidden_dim, mlp_proj_bias));
+            } else {
+                blocks["img_mlp"] = std::shared_ptr<GGMLBlock>(new MLP(hidden_size, mlp_hidden_dim, use_mlp_silu_act, mlp_proj_bias));
+            }

            if (!prune_mod && !share_modulation) {
                blocks["txt_mod"] = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
@@ -233,9 +282,11 @@ namespace Flux {
            blocks["txt_attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias, mlp_proj_bias));

            blocks["txt_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
-            blocks["txt_mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, mlp_hidden_dim * mlp_mult_factor, mlp_proj_bias));
-            // img_mlp.1 is nn.GELU(approximate="tanh")
-            blocks["txt_mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(mlp_hidden_dim, hidden_size, mlp_proj_bias));
+            if (use_yak_mlp) {
+                blocks["txt_mlp"] = std::shared_ptr<GGMLBlock>(new YakMLP(hidden_size, mlp_hidden_dim, mlp_proj_bias));
+            } else {
+                blocks["txt_mlp"] = std::shared_ptr<GGMLBlock>(new MLP(hidden_size, mlp_hidden_dim, use_mlp_silu_act, mlp_proj_bias));
+            }
        }

        std::vector<ModulationOut> get_distil_img_mod(GGMLRunnerContext* ctx, struct ggml_tensor* vec) {
@@ -272,15 +323,13 @@ namespace Flux {
            auto img_attn  = std::dynamic_pointer_cast<SelfAttention>(blocks["img_attn"]);

            auto img_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm2"]);
-            auto img_mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.0"]);
-            auto img_mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.2"]);
+            auto img_mlp   = std::dynamic_pointer_cast<UnaryBlock>(blocks["img_mlp"]);

            auto txt_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm1"]);
            auto txt_attn  = std::dynamic_pointer_cast<SelfAttention>(blocks["txt_attn"]);

            auto txt_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm2"]);
-            auto txt_mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["txt_mlp.0"]);
-            auto txt_mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["txt_mlp.2"]);
+            auto txt_mlp   = std::dynamic_pointer_cast<UnaryBlock>(blocks["txt_mlp"]);

            if (img_mods.empty()) {
                if (prune_mod) {
@@ -348,27 +397,15 @@ namespace Flux {
            // calculate the img bloks
            img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_attn->post_attention(ctx, img_attn_out), img_mod1.gate));

-            auto img_mlp_out = img_mlp_0->forward(ctx, Flux::modulate(ctx->ggml_ctx, img_norm2->forward(ctx, img), img_mod2.shift, img_mod2.scale));
-            if (use_mlp_silu_act) {
-                img_mlp_out = ggml_ext_silu_act(ctx->ggml_ctx, img_mlp_out);
-            } else {
-                img_mlp_out = ggml_gelu_inplace(ctx->ggml_ctx, img_mlp_out);
-            }
-            img_mlp_out = img_mlp_2->forward(ctx, img_mlp_out);
+            auto img_mlp_out = img_mlp->forward(ctx, Flux::modulate(ctx->ggml_ctx, img_norm2->forward(ctx, img), img_mod2.shift, img_mod2.scale));

            img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_mlp_out, img_mod2.gate));

            // calculate the txt bloks
            txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_attn->post_attention(ctx, txt_attn_out), txt_mod1.gate));

-            auto txt_mlp_out = txt_mlp_0->forward(ctx, Flux::modulate(ctx->ggml_ctx, txt_norm2->forward(ctx, txt), txt_mod2.shift, txt_mod2.scale));
-            if (use_mlp_silu_act) {
-                txt_mlp_out = ggml_ext_silu_act(ctx->ggml_ctx, txt_mlp_out);
-            } else {
-                txt_mlp_out = ggml_gelu_inplace(ctx->ggml_ctx, txt_mlp_out);
-            }
-            txt_mlp_out = txt_mlp_2->forward(ctx, txt_mlp_out);
-            txt         = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_mlp_out, txt_mod2.gate));
+            auto txt_mlp_out = txt_mlp->forward(ctx, Flux::modulate(ctx->ggml_ctx, txt_norm2->forward(ctx, txt), txt_mod2.shift, txt_mod2.scale));
+            txt              = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_mlp_out, txt_mod2.gate));

            return {img, txt};
        }
@@ -381,6 +418,7 @@ namespace Flux {
        int64_t mlp_hidden_dim;
        bool prune_mod;
        int idx = 0;
+        bool use_yak_mlp;
        bool use_mlp_silu_act;
        int64_t mlp_mult_factor;

@@ -393,8 +431,9 @@ namespace Flux {
                          bool prune_mod        = false,
                          bool share_modulation = false,
                          bool mlp_proj_bias    = true,
+                          bool use_yak_mlp      = false,
                          bool use_mlp_silu_act = false)
-            : hidden_size(hidden_size), num_heads(num_heads), idx(idx), prune_mod(prune_mod), use_mlp_silu_act(use_mlp_silu_act) {
+            : hidden_size(hidden_size), num_heads(num_heads), idx(idx), prune_mod(prune_mod), use_yak_mlp(use_yak_mlp), use_mlp_silu_act(use_mlp_silu_act) {
            int64_t head_dim = hidden_size / num_heads;
            float scale      = qk_scale;
            if (scale <= 0.f) {
@@ -402,7 +441,7 @@ namespace Flux {
            }
            mlp_hidden_dim  = hidden_size * mlp_ratio;
            mlp_mult_factor = 1;
-            if (use_mlp_silu_act) {
+            if (use_yak_mlp || use_mlp_silu_act) {
                mlp_mult_factor = 2;
            }

@@ -481,7 +520,9 @@ namespace Flux {
            k                = norm->key_norm(ctx, k);
            auto attn        = Rope::attention(ctx, q, k, v, pe, mask);  // [N, n_token, hidden_size]

-            if (use_mlp_silu_act) {
+            if (use_yak_mlp) {
+                mlp = ggml_ext_silu_act(ctx->ggml_ctx, mlp, false);
+            } else if (use_mlp_silu_act) {
                mlp = ggml_ext_silu_act(ctx->ggml_ctx, mlp);
            } else {
                mlp = ggml_gelu_inplace(ctx->ggml_ctx, mlp);
@@ -726,6 +767,8 @@ namespace Flux {
        int64_t in_dim              = 64;
        bool disable_bias           = false;
        bool share_modulation       = false;
+        bool semantic_txt_norm      = false;
+        bool use_yak_mlp            = false;
        bool use_mlp_silu_act       = false;
        float ref_index_scale       = 1.f;
        ChromaRadianceParams chroma_radiance_params;
@@ -759,6 +802,9 @@ namespace Flux {
                    blocks["guidance_in"] = std::make_shared<MLPEmbedder>(256, params.hidden_size, !params.disable_bias);
                }
            }
+            if (params.semantic_txt_norm) {
+                blocks["txt_norm"] = std::make_shared<RMSNorm>(params.context_in_dim);
+            }
            blocks["txt_in"] = std::make_shared<Linear>(params.context_in_dim, params.hidden_size, !params.disable_bias);

            for (int i = 0; i < params.depth; i++) {
@@ -770,6 +816,7 @@ namespace Flux {
                                                                                                   params.is_chroma,
                                                                                                   params.share_modulation,
                                                                                                   !params.disable_bias,
+                                                                                                   params.use_yak_mlp,
                                                                                                   params.use_mlp_silu_act);
            }

@@ -782,6 +829,7 @@ namespace Flux {
                                                                                                   params.is_chroma,
                                                                                                   params.share_modulation,
                                                                                                   !params.disable_bias,
+                                                                                                   params.use_yak_mlp,
                                                                                                   params.use_mlp_silu_act);
            }

@@ -948,6 +996,12 @@ namespace Flux {
                ss_mods     = single_stream_modulation->forward(ctx, vec);
            }

+            if (params.semantic_txt_norm) {
+                auto semantic_txt_norm = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm"]);
+
+                txt = semantic_txt_norm->forward(ctx, txt);
+            }
+
            txt = txt_in->forward(ctx, txt);

            for (int i = 0; i < params.depth; i++) {
@@ -1206,6 +1260,11 @@ namespace Flux {
            } else if (version == VERSION_CHROMA_RADIANCE) {
                flux_params.in_channels = 3;
                flux_params.patch_size  = 16;
+            } else if (version == VERSION_OVIS_IMAGE) {
+                flux_params.semantic_txt_norm = true;
+                flux_params.use_yak_mlp       = true;
+                flux_params.context_in_dim    = 2048;
+                flux_params.vec_in_dim        = 0;
            } else if (sd_version_is_flux2(version)) {
                flux_params.context_in_dim   = 15360;
                flux_params.in_channels      = 128;
@@ -1364,13 +1423,22 @@ namespace Flux {
                ref_latents[i] = to_backend(ref_latents[i]);
            }

+            std::set<int> txt_arange_dims;
+            if (sd_version_is_flux2(version)) {
+                txt_arange_dims    = {3};
+                increase_ref_index = true;
+            } else if (version == VERSION_OVIS_IMAGE) {
+                txt_arange_dims = {1, 2};
+            }
+
            pe_vec      = Rope::gen_flux_pe(x->ne[1],
                                            x->ne[0],
                                            flux_params.patch_size,
                                            x->ne[3],
                                            context->ne[1],
+                                            txt_arange_dims,
                                            ref_latents,
-                                       sd_version_is_flux2(version) ? true : increase_ref_index,
+                                            increase_ref_index,
                                            flux_params.ref_index_scale,
                                            flux_params.theta,
                                            flux_params.axes_dim);
@@ -1,4 +1,4 @@
-for f in *.cpp *.h *.hpp examples/cli/*.cpp examples/cli/*.h; do
+for f in *.cpp *.h *.hpp examples/cli/*.cpp examples/common/*.hpp examples/cli/*.h examples/server/*.cpp; do
  [[ "$f" == vocab* ]] && continue
  echo "formatting '$f'"
  # if [ "$f" != "stable-diffusion.h" ]; then
@@ -60,6 +60,14 @@
 #define SD_UNUSED(x) (void)(x)
 #endif

+__STATIC_INLINE__ int align_up_offset(int n, int multiple) {
+    return (multiple - n % multiple) % multiple;
+}
+
+__STATIC_INLINE__ int align_up(int n, int multiple) {
+    return n + align_up_offset(n, multiple);
+}
+
 __STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void*) {
    switch (level) {
        case GGML_LOG_LEVEL_DEBUG:
@@ -724,34 +732,22 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_slice(struct ggml_context* ctx,
 __STATIC_INLINE__ std::vector<struct ggml_tensor*> ggml_ext_chunk(struct ggml_context* ctx,
                                                                  struct ggml_tensor* x,
                                                                  int num,
-                                                                  int64_t dim) {
+                                                                  int64_t dim,
+                                                                  bool cont = true) {
    GGML_ASSERT(dim >= 0 && dim < 4);
    GGML_ASSERT(x->ne[dim] % num == 0);

-    int perm[4] = {0, 1, 2, 3};
-    for (int i = dim; i < 3; ++i)
-        perm[i] = perm[i + 1];
-    perm[3] = dim;
-
-    int inv_perm[4];
-    for (int i = 0; i < 4; ++i)
-        inv_perm[perm[i]] = i;
-
-    if (dim != 3) {
-        x = ggml_ext_torch_permute(ctx, x, perm[0], perm[1], perm[2], perm[3]);
-        x = ggml_cont(ctx, x);
-    }
-
    std::vector<struct ggml_tensor*> chunks;
-    int64_t chunk_size = x->ne[3] / num;
+    int64_t chunk_size  = x->ne[dim] / num;
+    int64_t stride      = chunk_size * x->nb[dim];
+    int64_t chunk_ne[4] = {x->ne[0], x->ne[1], x->ne[2], x->ne[3]};
+    chunk_ne[dim]       = chunk_size;
    for (int i = 0; i < num; i++) {
        auto chunk = ggml_view_4d(
            ctx, x,
-            x->ne[0], x->ne[1], x->ne[2], chunk_size,
-            x->nb[1], x->nb[2], x->nb[3], x->nb[3] * i * chunk_size);
-
-        if (dim != 3) {
-            chunk = ggml_ext_torch_permute(ctx, chunk, inv_perm[0], inv_perm[1], inv_perm[2], inv_perm[3]);
+            chunk_ne[0], chunk_ne[1], chunk_ne[2], chunk_ne[3],
+            x->nb[1], x->nb[2], x->nb[3], stride * i);
+        if (cont) {
            chunk = ggml_cont(ctx, chunk);
        }
        chunks.push_back(chunk);
@@ -760,17 +756,23 @@ __STATIC_INLINE__ std::vector<struct ggml_tensor*> ggml_ext_chunk(struct ggml_co
    return chunks;
 }

-__STATIC_INLINE__ ggml_tensor* ggml_ext_silu_act(ggml_context* ctx, ggml_tensor* x) {
+__STATIC_INLINE__ ggml_tensor* ggml_ext_silu_act(ggml_context* ctx, ggml_tensor* x, bool gate_first = true) {
    // x: [ne3, ne2, ne1, ne0]
    // return: [ne3, ne2, ne1, ne0/2]

-    auto x_vec = ggml_ext_chunk(ctx, x, 2, 0);
-    auto x1    = x_vec[0];  // [ne3, ne2, ne1, ne0/2]
-    auto x2    = x_vec[1];  // [ne3, ne2, ne1, ne0/2]
+    auto x_vec = ggml_ext_chunk(ctx, x, 2, 0, false);
+    ggml_tensor* gate;
+    if (gate_first) {
+        gate = x_vec[0];
+        x    = x_vec[1];
+    } else {
+        x    = x_vec[0];
+        gate = x_vec[1];
+    }
+    gate = ggml_cont(ctx, gate);
+    gate = ggml_silu_inplace(ctx, gate);

-    x1 = ggml_silu_inplace(ctx, x1);
-
-    x = ggml_mul(ctx, x1, x2);  // [ne3, ne2, ne1, ne0/2]
+    x = ggml_mul(ctx, x, gate);  // [ne3, ne2, ne1, ne0/2]

    return x;
 }
@@ -1386,10 +1388,14 @@ __STATIC_INLINE__ void ggml_ext_backend_tensor_get_and_sync(ggml_backend_t backe
 }

 __STATIC_INLINE__ float ggml_ext_backend_tensor_get_f32(ggml_tensor* tensor) {
-    GGML_ASSERT(tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_I32);
+    GGML_ASSERT(tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_I32 || tensor->type == GGML_TYPE_BF16);
    float value;
    if (tensor->type == GGML_TYPE_F32) {
        ggml_backend_tensor_get(tensor, &value, 0, sizeof(value));
+    } else if (tensor->type == GGML_TYPE_BF16) {
+        ggml_bf16_t bf16_value;
+        ggml_backend_tensor_get(tensor, &bf16_value, 0, sizeof(bf16_value));
+        value = ggml_bf16_to_fp32(bf16_value);
    } else if (tensor->type == GGML_TYPE_F16) {
        ggml_fp16_t f16_value;
        ggml_backend_tensor_get(tensor, &f16_value, 0, sizeof(f16_value));
@@ -91,6 +91,41 @@ const float flux_latent_rgb_proj[16][3] = {
    {-0.111849f, -0.055589f, -0.032361f}};
 float flux_latent_rgb_bias[3] = {0.024600f, -0.006937f, -0.008089f};

+const float flux2_latent_rgb_proj[32][3] = {
+    {0.000736f, -0.008385f, -0.019710f},
+    {-0.001352f, -0.016392f, 0.020693f},
+    {-0.006376f, 0.002428f, 0.036736f},
+    {0.039384f, 0.074167f, 0.119789f},
+    {0.007464f, -0.005705f, -0.004734f},
+    {-0.004086f, 0.005287f, -0.000409f},
+    {-0.032835f, 0.050802f, -0.028120f},
+    {-0.003158f, -0.000835f, 0.000406f},
+    {-0.112840f, -0.084337f, -0.023083f},
+    {0.001462f, -0.006656f, 0.000549f},
+    {-0.009980f, -0.007480f, 0.009702f},
+    {0.032540f, 0.000214f, -0.061388f},
+    {0.011023f, 0.000694f, 0.007143f},
+    {-0.001468f, -0.006723f, -0.001678f},
+    {-0.005921f, -0.010320f, -0.003907f},
+    {-0.028434f, 0.027584f, 0.018457f},
+    {0.014349f, 0.011523f, 0.000441f},
+    {0.009874f, 0.003081f, 0.001507f},
+    {0.002218f, 0.005712f, 0.001563f},
+    {0.053010f, -0.019844f, 0.008683f},
+    {-0.002507f, 0.005384f, 0.000938f},
+    {-0.002177f, -0.011366f, 0.003559f},
+    {-0.000261f, 0.015121f, -0.003240f},
+    {-0.003944f, -0.002083f, 0.005043f},
+    {-0.009138f, 0.011336f, 0.003781f},
+    {0.011429f, 0.003985f, -0.003855f},
+    {0.010518f, -0.005586f, 0.010131f},
+    {0.007883f, 0.002912f, -0.001473f},
+    {-0.003318f, -0.003160f, 0.003684f},
+    {-0.034560f, -0.008740f, 0.012996f},
+    {0.000166f, 0.001079f, -0.012153f},
+    {0.017772f, 0.000937f, -0.011953f}};
+float flux2_latent_rgb_bias[3] = {-0.028738f, -0.098463f, -0.107619f};
+
 // This one was taken straight from
 // https://github.com/Stability-AI/sd3.5/blob/8565799a3b41eb0c7ba976d18375f0f753f56402/sd3_impls.py#L288-L303
 // (MiT Licence)
@@ -128,16 +163,42 @@ const float sd_latent_rgb_proj[4][3] = {
    {-0.178022f, -0.200862f, -0.678514f}};
 float sd_latent_rgb_bias[3] = {-0.017478f, -0.055834f, -0.105825f};

-void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int width, int height, int frames, int dim) {
+void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
    size_t buffer_head = 0;
+
+    uint32_t latent_width  = latents->ne[0];
+    uint32_t latent_height = latents->ne[1];
+    uint32_t dim           = latents->ne[ggml_n_dims(latents) - 1];
+    uint32_t frames        = 1;
+    if (ggml_n_dims(latents) == 4) {
+        frames = latents->ne[2];
+    }
+
+    uint32_t rgb_width  = latent_width * patch_size;
+    uint32_t rgb_height = latent_height * patch_size;
+
+    uint32_t unpatched_dim = dim / (patch_size * patch_size);
+
    for (int k = 0; k < frames; k++) {
-        for (int j = 0; j < height; j++) {
-            for (int i = 0; i < width; i++) {
-                size_t latent_id = (i * latents->nb[0] + j * latents->nb[1] + k * latents->nb[2]);
+        for (int rgb_x = 0; rgb_x < rgb_width; rgb_x++) {
+            for (int rgb_y = 0; rgb_y < rgb_height; rgb_y++) {
+                int latent_x = rgb_x / patch_size;
+                int latent_y = rgb_y / patch_size;
+
+                int channel_offset = 0;
+                if (patch_size > 1) {
+                    channel_offset = ((rgb_y % patch_size) * patch_size + (rgb_x % patch_size));
+                }
+
+                size_t latent_id = (latent_x * latents->nb[0] + latent_y * latents->nb[1] + k * latents->nb[2]);
+
+                // should be incremented by 1 for each pixel
+                size_t pixel_id = k * rgb_width * rgb_height + rgb_y * rgb_width + rgb_x;
+
                float r = 0, g = 0, b = 0;
                if (latent_rgb_proj != nullptr) {
-                    for (int d = 0; d < dim; d++) {
-                        float value = *(float*)((char*)latents->data + latent_id + d * latents->nb[ggml_n_dims(latents) - 1]);
+                    for (int d = 0; d < unpatched_dim; d++) {
+                        float value = *(float*)((char*)latents->data + latent_id + (d * patch_size * patch_size + channel_offset) * latents->nb[ggml_n_dims(latents) - 1]);
                        r += value * latent_rgb_proj[d][0];
                        g += value * latent_rgb_proj[d][1];
                        b += value * latent_rgb_proj[d][2];
@@ -164,9 +225,9 @@ void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const fl
                g = g >= 0 ? g <= 1 ? g : 1 : 0;
                b = b >= 0 ? b <= 1 ? b : 1 : 0;

-                buffer[buffer_head++] = (uint8_t)(r * 255);
-                buffer[buffer_head++] = (uint8_t)(g * 255);
-                buffer[buffer_head++] = (uint8_t)(b * 255);
+                buffer[pixel_id * 3 + 0] = (uint8_t)(r * 255);
+                buffer[pixel_id * 3 + 1] = (uint8_t)(g * 255);
+                buffer[pixel_id * 3 + 2] = (uint8_t)(b * 255);
            }
        }
    }
@@ -356,6 +356,10 @@ namespace LLM {
                "<|fim_pad|>",
                "<|repo_name|>",
                "<|file_sep|>",
+                "<tool_response>",
+                "</tool_response>",
+                "<think>",
+                "</think>",
            };

            if (merges_utf8_str.size() > 0) {
@@ -859,11 +863,11 @@ namespace LLM {
            }

            if (arch == LLMArch::MISTRAL_SMALL_3_2) {
-                q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NORMAL, 131072, 1000000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
-                k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NORMAL, 131072, 1000000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
+                q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NORMAL, 8192, 1000000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
+                k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NORMAL, 8192, 1000000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
            } else if (arch == LLMArch::QWEN3) {
-                q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 151936, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
-                k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 151936, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
+                q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 40960, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
+                k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 40960, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
            } else {
                int sections[4] = {16, 24, 24, 0};
                q               = ggml_rope_multi(ctx->ggml_ctx, q, input_pos, nullptr, head_dim, sections, GGML_ROPE_TYPE_MROPE, 128000, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
@@ -1073,29 +1077,22 @@ namespace LLM {
            : GGMLRunner(backend, offload_params_to_cpu), enable_vision(enable_vision_) {
            params.arch = arch;
            if (arch == LLMArch::MISTRAL_SMALL_3_2) {
-                params.num_layers        = 40;
-                params.hidden_size       = 5120;
-                params.intermediate_size = 32768;
-                params.head_dim          = 128;
-                params.num_heads         = 32;
-                params.num_kv_heads      = 8;
-                params.qkv_bias          = false;
-                params.vocab_size        = 131072;
-                params.rms_norm_eps      = 1e-5f;
+                params.head_dim     = 128;
+                params.num_heads    = 32;
+                params.num_kv_heads = 8;
+                params.qkv_bias     = false;
+                params.rms_norm_eps = 1e-5f;
            } else if (arch == LLMArch::QWEN3) {
-                params.num_layers        = 36;
-                params.hidden_size       = 2560;
-                params.intermediate_size = 9728;
-                params.head_dim          = 128;
-                params.num_heads         = 32;
-                params.num_kv_heads      = 8;
-                params.qkv_bias          = false;
-                params.qk_norm           = true;
-                params.vocab_size        = 151936;
-                params.rms_norm_eps      = 1e-6f;
+                params.head_dim     = 128;
+                params.num_heads    = 32;
+                params.num_kv_heads = 8;
+                params.qkv_bias     = false;
+                params.qk_norm      = true;
+                params.rms_norm_eps = 1e-6f;
            }
            bool have_vision_weight = false;
            bool llama_cpp_style    = false;
+            params.num_layers       = 0;
            for (auto pair : tensor_storage_map) {
                std::string tensor_name = pair.first;
                if (tensor_name.find(prefix) == std::string::npos)
@@ -1105,10 +1102,36 @@ namespace LLM {
                    have_vision_weight = true;
                    if (contains(tensor_name, "attn.q_proj")) {
                        llama_cpp_style = true;
-                        break;
+                    }
+                    continue;
+                }
+                pos = tensor_name.find("layers.");
+                if (pos != std::string::npos) {
+                    tensor_name = tensor_name.substr(pos);  // remove prefix
+                    auto items  = split_string(tensor_name, '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        if (block_index + 1 > params.num_layers) {
+                            params.num_layers = block_index + 1;
+                        }
                    }
                }
+                if (contains(tensor_name, "embed_tokens.weight")) {
+                    params.hidden_size = pair.second.ne[0];
+                    params.vocab_size  = pair.second.ne[1];
+                }
+                if (contains(tensor_name, "layers.0.mlp.gate_proj.weight")) {
+                    params.intermediate_size = pair.second.ne[1];
+                }
            }
+            if (arch == LLMArch::QWEN3 && params.num_layers == 28) {  // Qwen3 2B
+                params.num_heads = 16;
+            }
+            LOG_DEBUG("llm: num_layers = %" PRId64 ", vocab_size = %" PRId64 ", hidden_size = %" PRId64 ", intermediate_size = %" PRId64,
+                      params.num_layers,
+                      params.vocab_size,
+                      params.hidden_size,
+                      params.intermediate_size);
            if (enable_vision && !have_vision_weight) {
                LOG_WARN("no vision weights detected, vision disabled");
                enable_vision = false;
@@ -1056,6 +1056,9 @@ SDVersion ModelLoader::get_sd_version() {
            if (tensor_storage.name.find("model.diffusion_model.double_stream_modulation_img.lin.weight") != std::string::npos) {
                return VERSION_FLUX2;
            }
+            if (tensor_storage.name.find("model.diffusion_model.double_blocks.0.img_mlp.gate_proj.weight") != std::string::npos) {
+                return VERSION_OVIS_IMAGE;
+            }
            if (tensor_storage.name.find("model.diffusion_model.cap_embedder.0.weight") != std::string::npos) {
                return VERSION_Z_IMAGE;
            }
@@ -45,6 +45,7 @@ enum SDVersion {
    VERSION_QWEN_IMAGE,
    VERSION_FLUX2,
    VERSION_Z_IMAGE,
+    VERSION_OVIS_IMAGE,
    VERSION_COUNT,
 };

@@ -90,6 +91,7 @@ static inline bool sd_version_is_flux(SDVersion version) {
        version == VERSION_FLUX_FILL ||
        version == VERSION_FLUX_CONTROLS ||
        version == VERSION_FLEX_2 ||
+        version == VERSION_OVIS_IMAGE ||
        version == VERSION_CHROMA_RADIANCE) {
        return true;
    }
@@ -72,11 +72,13 @@ namespace Rope {
    }

    // Generate IDs for image patches and text
-    __STATIC_INLINE__ std::vector<std::vector<float>> gen_flux_txt_ids(int bs, int context_len, int axes_dim_num) {
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_flux_txt_ids(int bs, int context_len, int axes_dim_num, std::set<int> arange_dims) {
        auto txt_ids = std::vector<std::vector<float>>(bs * context_len, std::vector<float>(axes_dim_num, 0.0f));
-        if (axes_dim_num == 4) {
-            for (int i = 0; i < bs * context_len; i++) {
-                txt_ids[i][3] = (i % context_len);
+        for (int dim = 0; dim < axes_dim_num; dim++) {
+            if (arange_dims.find(dim) != arange_dims.end()) {
+                for (int i = 0; i < bs * context_len; i++) {
+                    txt_ids[i][dim] = (i % context_len);
+                }
            }
        }
        return txt_ids;
@@ -211,10 +213,11 @@ namespace Rope {
                                                                   int bs,
                                                                   int axes_dim_num,
                                                                   int context_len,
+                                                                   std::set<int> txt_arange_dims,
                                                                   const std::vector<ggml_tensor*>& ref_latents,
                                                                   bool increase_ref_index,
                                                                   float ref_index_scale) {
-        auto txt_ids = gen_flux_txt_ids(bs, context_len, axes_dim_num);
+        auto txt_ids = gen_flux_txt_ids(bs, context_len, axes_dim_num, txt_arange_dims);
        auto img_ids = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num);

        auto ids = concat_ids(txt_ids, img_ids, bs);
@@ -231,6 +234,7 @@ namespace Rope {
                                                     int patch_size,
                                                     int bs,
                                                     int context_len,
+                                                     std::set<int> txt_arange_dims,
                                                     const std::vector<ggml_tensor*>& ref_latents,
                                                     bool increase_ref_index,
                                                     float ref_index_scale,
@@ -242,6 +246,7 @@ namespace Rope {
                                                           bs,
                                                           static_cast<int>(axes_dim.size()),
                                                           context_len,
+                                                           txt_arange_dims,
                                                           ref_latents,
                                                           increase_ref_index,
                                                           ref_index_scale);
@@ -46,6 +46,7 @@ const char* model_version_to_str[] = {
    "Qwen Image",
    "Flux.2",
    "Z-Image",
+    "Ovis Image",
 };

 const char* sampling_methods_str[] = {
@@ -424,6 +425,13 @@ public:
                                                                        tensor_storage_map,
                                                                        sd_ctx_params->chroma_use_t5_mask,
                                                                        sd_ctx_params->chroma_t5_mask_pad);
+                } else if (version == VERSION_OVIS_IMAGE) {
+                    cond_stage_model = std::make_shared<LLMEmbedder>(clip_backend,
+                                                                     offload_params_to_cpu,
+                                                                     tensor_storage_map,
+                                                                     version,
+                                                                     "",
+                                                                     false);
                } else {
                    cond_stage_model = std::make_shared<FluxCLIPEmbedder>(clip_backend,
                                                                          offload_params_to_cpu,
@@ -500,18 +508,22 @@ public:
                                                                "model.diffusion_model",
                                                                version);
            } else {  // SD1.x SD2.x SDXL
+                std::map<std::string, std::string> embbeding_map;
+                for (int i = 0; i < sd_ctx_params->embedding_count; i++) {
+                    embbeding_map.emplace(SAFE_STR(sd_ctx_params->embeddings[i].name), SAFE_STR(sd_ctx_params->embeddings[i].path));
+                }
                if (strstr(SAFE_STR(sd_ctx_params->photo_maker_path), "v2")) {
                    cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend,
                                                                                           offload_params_to_cpu,
                                                                                           tensor_storage_map,
-                                                                                           SAFE_STR(sd_ctx_params->embedding_dir),
+                                                                                           embbeding_map,
                                                                                           version,
                                                                                           PM_VERSION_2);
                } else {
                    cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend,
                                                                                           offload_params_to_cpu,
                                                                                           tensor_storage_map,
-                                                                                           SAFE_STR(sd_ctx_params->embedding_dir),
+                                                                                           embbeding_map,
                                                                                           version);
                }
                diffusion_model = std::make_shared<UNetModel>(backend,
@@ -690,6 +702,11 @@ public:
            ignore_tensors.insert("first_stage_model.quant");
            ignore_tensors.insert("text_encoders.llm.visual.");
        }
+        if (version == VERSION_OVIS_IMAGE) {
+            ignore_tensors.insert("text_encoders.llm.vision_model.");
+            ignore_tensors.insert("text_encoders.llm.visual_tokenizer.");
+            ignore_tensors.insert("text_encoders.llm.vte.");
+        }
        if (version == VERSION_SVD) {
            ignore_tensors.insert("conditioner.embedders.3");
        }
@@ -920,28 +937,17 @@ public:
                                                         float multiplier,
                                                         ggml_backend_t backend,
                                                         LoraModel::filter_t lora_tensor_filter = nullptr) {
-        std::string lora_name      = lora_id;
-        std::string high_noise_tag = "|high_noise|";
-        bool is_high_noise         = false;
-        if (starts_with(lora_name, high_noise_tag)) {
-            lora_name     = lora_name.substr(high_noise_tag.size());
+        std::string lora_path             = lora_id;
+        static std::string high_noise_tag = "|high_noise|";
+        bool is_high_noise                = false;
+        if (starts_with(lora_path, high_noise_tag)) {
+            lora_path     = lora_path.substr(high_noise_tag.size());
            is_high_noise = true;
-            LOG_DEBUG("high noise lora: %s", lora_name.c_str());
+            LOG_DEBUG("high noise lora: %s", lora_path.c_str());
        }
-        std::string st_file_path   = path_join(lora_model_dir, lora_name + ".safetensors");
-        std::string ckpt_file_path = path_join(lora_model_dir, lora_name + ".ckpt");
-        std::string file_path;
-        if (file_exists(st_file_path)) {
-            file_path = st_file_path;
-        } else if (file_exists(ckpt_file_path)) {
-            file_path = ckpt_file_path;
-        } else {
-            LOG_WARN("can not find %s or %s for lora %s", st_file_path.c_str(), ckpt_file_path.c_str(), lora_name.c_str());
-            return nullptr;
-        }
-        auto lora = std::make_shared<LoraModel>(lora_id, backend, file_path, is_high_noise ? "model.high_noise_" : "", version);
+        auto lora = std::make_shared<LoraModel>(lora_id, backend, lora_path, is_high_noise ? "model.high_noise_" : "", version);
        if (!lora->load_from_file(n_threads, lora_tensor_filter)) {
-            LOG_WARN("load lora tensors from %s failed", file_path.c_str());
+            LOG_WARN("load lora tensors from %s failed", lora_path.c_str());
            return nullptr;
        }

@@ -1126,12 +1132,15 @@ public:
        }
    }

-    std::string apply_loras_from_prompt(const std::string& prompt) {
-        auto result_pair                                = extract_and_remove_lora(prompt);
-        std::unordered_map<std::string, float> lora_f2m = result_pair.first;  // lora_name -> multiplier
-
-        for (auto& kv : lora_f2m) {
-            LOG_DEBUG("lora %s:%.2f", kv.first.c_str(), kv.second);
+    void apply_loras(const sd_lora_t* loras, uint32_t lora_count) {
+        std::unordered_map<std::string, float> lora_f2m;
+        for (int i = 0; i < lora_count; i++) {
+            std::string lora_id = SAFE_STR(loras[i].path);
+            if (loras[i].is_high_noise) {
+                lora_id = "|high_noise|" + lora_id;
+            }
+            lora_f2m[lora_id] = loras[i].multiplier;
+            LOG_DEBUG("lora %s:%.2f", lora_id.c_str(), loras[i].multiplier);
        }
        int64_t t0 = ggml_time_ms();
        if (apply_lora_immediately) {
@@ -1142,9 +1151,7 @@ public:
        int64_t t1 = ggml_time_ms();
        if (!lora_f2m.empty()) {
            LOG_INFO("apply_loras completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
-            LOG_DEBUG("prompt after extract and remove lora: \"%s\"", result_pair.second.c_str());
        }
-        return result_pair.second;
    }

    ggml_tensor* id_encoder(ggml_context* work_ctx,
@@ -1309,10 +1316,17 @@ public:
        uint32_t dim           = latents->ne[ggml_n_dims(latents) - 1];

        if (preview_mode == PREVIEW_PROJ) {
+            int64_t patch_sz                       = 1;
            const float(*latent_rgb_proj)[channel] = nullptr;
            float* latent_rgb_bias                 = nullptr;

-            if (dim == 48) {
+            if (dim == 128) {
+                if (sd_version_is_flux2(version)) {
+                    latent_rgb_proj = flux2_latent_rgb_proj;
+                    latent_rgb_bias = flux2_latent_rgb_bias;
+                    patch_sz        = 2;
+                }
+            } else if (dim == 48) {
                if (sd_version_is_wan(version)) {
                    latent_rgb_proj = wan_22_latent_rgb_proj;
                    latent_rgb_bias = wan_22_latent_rgb_bias;
@@ -1365,12 +1379,15 @@ public:
                frames = latents->ne[2];
            }

-            uint8_t* data = (uint8_t*)malloc(frames * width * height * channel * sizeof(uint8_t));
+            uint32_t img_width  = width * patch_sz;
+            uint32_t img_height = height * patch_sz;

-            preview_latent_video(data, latents, latent_rgb_proj, latent_rgb_bias, width, height, frames, dim);
+            uint8_t* data = (uint8_t*)malloc(frames * img_width * img_height * channel * sizeof(uint8_t));
+
+            preview_latent_video(data, latents, latent_rgb_proj, latent_rgb_bias, patch_sz);
            sd_image_t* images = (sd_image_t*)malloc(frames * sizeof(sd_image_t));
            for (int i = 0; i < frames; i++) {
-                images[i] = {width, height, channel, data + i * width * height * channel};
+                images[i] = {img_width, img_height, channel, data + i * img_width * img_height * channel};
            }
            step_callback(step, frames, images, is_noisy, step_callback_data);
            free(data);
@@ -1881,6 +1898,18 @@ public:
        return vae_scale_factor;
    }

+    int get_diffusion_model_down_factor() {
+        int down_factor = 8;  // unet
+        if (sd_version_is_dit(version)) {
+            if (sd_version_is_wan(version)) {
+                down_factor = 2;
+            } else {
+                down_factor = 1;
+            }
+        }
+        return down_factor;
+    }
+
    int get_latent_channel() {
        int latent_channel = 4;
        if (sd_version_is_dit(version)) {
@@ -2508,7 +2537,6 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
             "taesd_path: %s\n"
             "control_net_path: %s\n"
             "lora_model_dir: %s\n"
-             "embedding_dir: %s\n"
             "photo_maker_path: %s\n"
             "tensor_type_rules: %s\n"
             "vae_decode_only: %s\n"
@@ -2539,7 +2567,6 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
             SAFE_STR(sd_ctx_params->taesd_path),
             SAFE_STR(sd_ctx_params->control_net_path),
             SAFE_STR(sd_ctx_params->lora_model_dir),
-             SAFE_STR(sd_ctx_params->embedding_dir),
             SAFE_STR(sd_ctx_params->photo_maker_path),
             SAFE_STR(sd_ctx_params->tensor_type_rules),
             BOOL_STR(sd_ctx_params->vae_decode_only),
@@ -2573,6 +2600,8 @@ void sd_sample_params_init(sd_sample_params_t* sample_params) {
    sample_params->scheduler                   = SCHEDULER_COUNT;
    sample_params->sample_method               = SAMPLE_METHOD_COUNT;
    sample_params->sample_steps                = 20;
+    sample_params->custom_sigmas               = nullptr;
+    sample_params->custom_sigmas_count         = 0;
 }

 char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
@@ -2790,8 +2819,6 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
    int sample_steps = sigmas.size() - 1;

    int64_t t0 = ggml_time_ms();
-    // Apply lora
-    prompt = sd_ctx->sd->apply_loras_from_prompt(prompt);

    // Photo Maker
    std::string prompt_text_only;
@@ -3120,22 +3147,19 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
    sd_ctx->sd->vae_tiling_params = sd_img_gen_params->vae_tiling_params;
    int width                     = sd_img_gen_params->width;
    int height                    = sd_img_gen_params->height;
-    int vae_scale_factor          = sd_ctx->sd->get_vae_scale_factor();
-    if (sd_version_is_dit(sd_ctx->sd->version)) {
-        if (width % 16 || height % 16) {
-            LOG_ERROR("Image dimensions must be must be a multiple of 16 on each axis for %s models. (Got %dx%d)",
-                      model_version_to_str[sd_ctx->sd->version],
-                      width,
-                      height);
-            return nullptr;
-        }
-    } else if (width % 64 || height % 64) {
-        LOG_ERROR("Image dimensions must be must be a multiple of 64 on each axis for %s models. (Got %dx%d)",
-                  model_version_to_str[sd_ctx->sd->version],
-                  width,
-                  height);
-        return nullptr;
+
+    int vae_scale_factor            = sd_ctx->sd->get_vae_scale_factor();
+    int diffusion_model_down_factor = sd_ctx->sd->get_diffusion_model_down_factor();
+    int spatial_multiple            = vae_scale_factor * diffusion_model_down_factor;
+
+    int width_offset  = align_up_offset(width, spatial_multiple);
+    int height_offset = align_up_offset(height, spatial_multiple);
+    if (width_offset > 0 || height_offset > 0) {
+        width += width_offset;
+        height += height_offset;
+        LOG_WARN("align up %dx%d to %dx%d (multiple=%d)", sd_img_gen_params->width, sd_img_gen_params->height, width, height, spatial_multiple);
    }
+
    LOG_DEBUG("generate_image %dx%d", width, height);
    if (sd_ctx == nullptr || sd_img_gen_params == nullptr) {
        return nullptr;
@@ -3163,17 +3187,30 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g

    size_t t0 = ggml_time_ms();

+    // Apply lora
+    sd_ctx->sd->apply_loras(sd_img_gen_params->loras, sd_img_gen_params->lora_count);
+
    enum sample_method_t sample_method = sd_img_gen_params->sample_params.sample_method;
    if (sample_method == SAMPLE_METHOD_COUNT) {
        sample_method = sd_get_default_sample_method(sd_ctx);
    }
    LOG_INFO("sampling using %s method", sampling_methods_str[sample_method]);

-    int sample_steps          = sd_img_gen_params->sample_params.sample_steps;
-    std::vector<float> sigmas = sd_ctx->sd->denoiser->get_sigmas(sample_steps,
-                                                                 sd_ctx->sd->get_image_seq_len(height, width),
-                                                                 sd_img_gen_params->sample_params.scheduler,
-                                                                 sd_ctx->sd->version);
+    int sample_steps = sd_img_gen_params->sample_params.sample_steps;
+    std::vector<float> sigmas;
+    if (sd_img_gen_params->sample_params.custom_sigmas_count > 0) {
+        sigmas = std::vector<float>(sd_img_gen_params->sample_params.custom_sigmas,
+                                    sd_img_gen_params->sample_params.custom_sigmas + sd_img_gen_params->sample_params.custom_sigmas_count);
+        if (sample_steps != sigmas.size() - 1) {
+            sample_steps = static_cast<int>(sigmas.size()) - 1;
+            LOG_WARN("sample_steps != custom_sigmas_count - 1, set sample_steps to %d", sample_steps);
+        }
+    } else {
+        sigmas = sd_ctx->sd->denoiser->get_sigmas(sample_steps,
+                                                  sd_ctx->sd->get_image_seq_len(height, width),
+                                                  sd_img_gen_params->sample_params.scheduler,
+                                                  sd_ctx->sd->version);
+    }

    ggml_tensor* init_latent   = nullptr;
    ggml_tensor* concat_latent = nullptr;
@@ -3406,9 +3443,19 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
    int frames       = sd_vid_gen_params->video_frames;
    frames           = (frames - 1) / 4 * 4 + 1;
    int sample_steps = sd_vid_gen_params->sample_params.sample_steps;
-    LOG_INFO("generate_video %dx%dx%d", width, height, frames);

-    int vae_scale_factor = sd_ctx->sd->get_vae_scale_factor();
+    int vae_scale_factor            = sd_ctx->sd->get_vae_scale_factor();
+    int diffusion_model_down_factor = sd_ctx->sd->get_diffusion_model_down_factor();
+    int spatial_multiple            = vae_scale_factor * diffusion_model_down_factor;
+
+    int width_offset  = align_up_offset(width, spatial_multiple);
+    int height_offset = align_up_offset(height, spatial_multiple);
+    if (width_offset > 0 || height_offset > 0) {
+        width += width_offset;
+        height += height_offset;
+        LOG_WARN("align up %dx%d to %dx%d (multiple=%d)", sd_vid_gen_params->width, sd_vid_gen_params->height, width, height, spatial_multiple);
+    }
+    LOG_INFO("generate_video %dx%dx%d", width, height, frames);

    enum sample_method_t sample_method = sd_vid_gen_params->sample_params.sample_method;
    if (sample_method == SAMPLE_METHOD_COUNT) {
@@ -3426,7 +3473,29 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
    if (high_noise_sample_steps > 0) {
        total_steps += high_noise_sample_steps;
    }
-    std::vector<float> sigmas = sd_ctx->sd->denoiser->get_sigmas(total_steps, 0, sd_vid_gen_params->sample_params.scheduler, sd_ctx->sd->version);
+
+    std::vector<float> sigmas;
+    if (sd_vid_gen_params->sample_params.custom_sigmas_count > 0) {
+        sigmas = std::vector<float>(sd_vid_gen_params->sample_params.custom_sigmas,
+                                    sd_vid_gen_params->sample_params.custom_sigmas + sd_vid_gen_params->sample_params.custom_sigmas_count);
+        if (total_steps != sigmas.size() - 1) {
+            total_steps = static_cast<int>(sigmas.size()) - 1;
+            LOG_WARN("total_steps != custom_sigmas_count - 1, set total_steps to %d", total_steps);
+            if (sample_steps >= total_steps) {
+                sample_steps = total_steps;
+                LOG_WARN("total_steps != custom_sigmas_count - 1, set sample_steps to %d", sample_steps);
+            }
+            if (high_noise_sample_steps > 0) {
+                high_noise_sample_steps = total_steps - sample_steps;
+                LOG_WARN("total_steps != custom_sigmas_count - 1, set high_noise_sample_steps to %d", high_noise_sample_steps);
+            }
+        }
+    } else {
+        sigmas = sd_ctx->sd->denoiser->get_sigmas(total_steps,
+                                                  0,
+                                                  sd_vid_gen_params->sample_params.scheduler,
+                                                  sd_ctx->sd->version);
+    }

    if (high_noise_sample_steps < 0) {
        // timesteps ∝ sigmas for Flow models (like wan2.2 a14b)
@@ -3462,7 +3531,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
    int64_t t0 = ggml_time_ms();

    // Apply lora
-    prompt = sd_ctx->sd->apply_loras_from_prompt(prompt);
+    sd_ctx->sd->apply_loras(sd_vid_gen_params->loras, sd_vid_gen_params->lora_count);

    ggml_tensor* init_latent        = nullptr;
    ggml_tensor* clip_vision_output = nullptr;
@@ -3806,4 +3875,4 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
    LOG_INFO("generate_video completed in %.2fs", (t5 - t0) * 1.0f / 1000);

    return result_images;
-}
+}
@@ -150,6 +150,11 @@ typedef struct {
    float rel_size_y;
 } sd_tiling_params_t;

+typedef struct {
+    const char* name;
+    const char* path;
+} sd_embedding_t;
+
 typedef struct {
    const char* model_path;
    const char* clip_l_path;
@@ -164,7 +169,8 @@ typedef struct {
    const char* taesd_path;
    const char* control_net_path;
    const char* lora_model_dir;
-    const char* embedding_dir;
+    const sd_embedding_t* embeddings;
+    uint32_t embedding_count;
    const char* photo_maker_path;
    const char* tensor_type_rules;
    bool vae_decode_only;
@@ -219,6 +225,8 @@ typedef struct {
    int sample_steps;
    float eta;
    int shifted_timestep;
+    float* custom_sigmas;
+    int custom_sigmas_count;
 } sd_sample_params_t;

 typedef struct {
@@ -236,6 +244,14 @@ typedef struct {
 } sd_easycache_params_t;

 typedef struct {
+    bool is_high_noise;
+    float multiplier;
+    const char* path;
+} sd_lora_t;
+
+typedef struct {
+    const sd_lora_t* loras;
+    uint32_t lora_count;
    const char* prompt;
    const char* negative_prompt;
    int clip_skip;
@@ -259,6 +275,8 @@ typedef struct {
 } sd_img_gen_params_t;

 typedef struct {
+    const sd_lora_t* loras;
+    uint32_t lora_count;
    const char* prompt;
    const char* negative_prompt;
    int clip_skip;
@@ -331,7 +349,8 @@ typedef struct upscaler_ctx_t upscaler_ctx_t;
 SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
                                        bool offload_params_to_cpu,
                                        bool direct,
-                                        int n_threads);
+                                        int n_threads,
+                                        int tile_size);
 SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);

 SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx,
@@ -353,6 +372,9 @@ SD_API bool preprocess_canny(sd_image_t image,
                             float strong,
                             bool inverse);

+SD_API const char* sd_commit(void);
+SD_API const char* sd_version(void);
+
 #ifdef __cplusplus
 }
 #endif
@@ -1,3 +1,10 @@
 - json.hpp library from: https://github.com/nlohmann/json
+    - LICENSE: https://github.com/nlohmann/json/blob/develop/LICENSE.MIT
 - ZIP Library from: https://github.com/kuba--/zip
- darts.h from: https://github.com/google/sentencepiece/tree/master/third_party/darts_clone
+    LICENSE: https://github.com/kuba--/zip/blob/master/LICENSE.txt
+- darts.h from: https://github.com/google/sentencepiece/tree/master/third_party/darts_clone
+    - LICENSE: https://github.com/google/sentencepiece/blob/master/third_party/darts_clone/LICENSE
+- httplib.h from: https://github.com/yhirose/cpp-httplib/blob/master/httplib.h
+    - LICENSE: https://github.com/yhirose/cpp-httplib/blob/master/LICENSE
+- stb_image.h/stb_image_resize.h/stb_image_write.h from: https://github.com/nothings/stb
+    - LICENSE: https://github.com/nothings/stb/blob/master/LICENSE
@@ -9,12 +9,15 @@ struct UpscalerGGML {
    std::shared_ptr<ESRGAN> esrgan_upscaler;
    std::string esrgan_path;
    int n_threads;
-    bool direct = false;
+    bool direct   = false;
+    int tile_size = 128;

    UpscalerGGML(int n_threads,
-                 bool direct = false)
+                 bool direct   = false,
+                 int tile_size = 128)
        : n_threads(n_threads),
-          direct(direct) {
+          direct(direct),
+          tile_size(tile_size) {
    }

    bool load_from_file(const std::string& esrgan_path,
@@ -51,7 +54,7 @@ struct UpscalerGGML {
            backend = ggml_backend_cpu_init();
        }
        LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
-        esrgan_upscaler = std::make_shared<ESRGAN>(backend, offload_params_to_cpu, model_loader.get_tensor_storage_map());
+        esrgan_upscaler = std::make_shared<ESRGAN>(backend, offload_params_to_cpu, tile_size, model_loader.get_tensor_storage_map());
        if (direct) {
            esrgan_upscaler->set_conv2d_direct_enabled(true);
        }
@@ -113,14 +116,15 @@ struct upscaler_ctx_t {
 upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
                                 bool offload_params_to_cpu,
                                 bool direct,
-                                 int n_threads) {
+                                 int n_threads,
+                                 int tile_size) {
    upscaler_ctx_t* upscaler_ctx = (upscaler_ctx_t*)malloc(sizeof(upscaler_ctx_t));
    if (upscaler_ctx == nullptr) {
        return nullptr;
    }
    std::string esrgan_path(esrgan_path_c_str);

-    upscaler_ctx->upscaler = new UpscalerGGML(n_threads, direct);
+    upscaler_ctx->upscaler = new UpscalerGGML(n_threads, direct, tile_size);
    if (upscaler_ctx->upscaler == nullptr) {
        return nullptr;
    }
@@ -95,20 +95,6 @@ bool is_directory(const std::string& path) {
    return (attributes != INVALID_FILE_ATTRIBUTES && (attributes & FILE_ATTRIBUTE_DIRECTORY));
 }

-std::string get_full_path(const std::string& dir, const std::string& filename) {
-    std::string full_path = dir + "\\" + filename;
-
-    WIN32_FIND_DATA find_file_data;
-    HANDLE hFind = FindFirstFile(full_path.c_str(), &find_file_data);
-
-    if (hFind != INVALID_HANDLE_VALUE) {
-        FindClose(hFind);
-        return full_path;
-    } else {
-        return "";
-    }
-}
-
 #else  // Unix
 #include <dirent.h>
 #include <sys/stat.h>
@@ -123,26 +109,6 @@ bool is_directory(const std::string& path) {
    return (stat(path.c_str(), &buffer) == 0 && S_ISDIR(buffer.st_mode));
 }

-// TODO: add windows version
-std::string get_full_path(const std::string& dir, const std::string& filename) {
-    DIR* dp = opendir(dir.c_str());
-
-    if (dp != nullptr) {
-        struct dirent* entry;
-
-        while ((entry = readdir(dp)) != nullptr) {
-            if (strcasecmp(entry->d_name, filename.c_str()) == 0) {
-                closedir(dp);
-                return dir + "/" + entry->d_name;
-            }
-        }
-
-        closedir(dp);
-    }
-
-    return "";
-}
-
 #endif

 // get_num_physical_cores is copy from
@@ -22,7 +22,6 @@ int round_up_to(int value, int base);

 bool file_exists(const std::string& filename);
 bool is_directory(const std::string& path);
-std::string get_full_path(const std::string& dir, const std::string& filename);

 std::u32string utf8_to_utf32(const std::string& utf8_str);
 std::string utf32_to_utf8(const std::u32string& utf32_str);
@@ -0,0 +1,20 @@
+#include "stable-diffusion.h"
+
+#ifndef SDCPP_BUILD_COMMIT
+#define SDCPP_BUILD_COMMIT unknown
+#endif
+
+#ifndef SDCPP_BUILD_VERSION
+#define SDCPP_BUILD_VERSION unknown
+#endif
+
+#define STRINGIZE2(x) #x
+#define STRINGIZE(x) STRINGIZE2(x)
+
+const char* sd_commit(void) {
+    return STRINGIZE(SDCPP_BUILD_COMMIT);
+}
+
+const char* sd_version(void) {
+    return STRINGIZE(SDCPP_BUILD_VERSION);
+}
Author	SHA1	Message	Date
stduhpf	d96b4152d6	perf: optimize ggml_ext_chunk (#1084 )	2025-12-14 01:22:41 +08:00
rmatifandleejet	8f05f5bc6e	feat: add support for custom scheduler (#694 ) --------- Co-authored-by: leejet <leejet714@gmail.com>	2025-12-13 16:20:02 +08:00
leejet	15d0f82760	feat(server): do not parse lora fromt client-side prompts (#1083 )	2025-12-13 14:27:47 +08:00
xxnuo	6888fcb581	feat: server add default_gen_params to override default args (#1050 )	2025-12-13 14:22:32 +08:00
leejet	2aecdd57ca	feat: simple openai image generation api compatiple server (#1037 )	2025-12-13 13:53:21 +08:00
leejet	11ab095230	fix: resolve embedding loading issue when calling generate_image multiple times (#1078 )	2025-12-12 23:08:12 +08:00
Wagner Bruna	a3a88fc9b2	fix: avoid crash loading LoRAs with bf16 weights (#1077 )	2025-12-12 22:36:54 +08:00
leejet	8823dc48bc	feat: align the spatial size to the corresponding multiple (#1073 )	2025-12-10 23:15:08 +08:00
Pedrito	1ac5a616de	feat: support custom upscale tile size (#896 )	2025-12-10 22:25:19 +08:00
leejet	d939f6e86a	refactor: optimize the handling of LoRA models (#1070 )	2025-12-10 00:26:07 +08:00
Wagner Bruna	e72aea796e	feat: embed version string and git commit hash (#1008 )	2025-12-09 22:38:54 +08:00
wuhei	a908436729	docs: update download link for Stable Diffusion v1.5 (#1063 )	2025-12-09 22:06:16 +08:00
stduhpf	583a02e29e	feat: add Flux.2 VAE proj matrix for previews (#1017 )	2025-12-09 22:00:45 +08:00
leejet	96c3e64057	refactor: optimize the handling of embedding (#1068 ) * optimize the handling of embedding * support case-insensitive embedding names	2025-12-08 23:59:04 +08:00
Weiqi Gao	0392273e10	chore: add compute kernels to Windows CUDA build (#1062 ) * Fix syntax for CUDA architecture definitions * Extend CUDA support to GTX 10 Series to RTX 50 Series * update cuda installer step version to install cuda 12.8.1 * Remove unsupported compute capability	2025-12-07 22:12:50 +08:00
leejet	bf1a388b44	docs: update logo	2025-12-07 15:09:32 +08:00
leejet	c9005337a8	docs: update logo	2025-12-07 14:56:21 +08:00
leejet	2f0bd31a84	feat: add ovis image support (#1057 )	2025-12-07 12:32:56 +08:00