refactor: move all cache parameter defaults to the library (#1327)

.github/workflows/build.yml

@@ -70,7 +70,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -123,7 +123,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -177,7 +177,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
@@ -240,7 +240,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -340,7 +340,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Pack artifacts
         id: pack_artifacts
@@ -463,7 +463,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Pack artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -581,7 +581,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Prepare artifacts
         id: prepare_artifacts
@@ -660,7 +660,7 @@ jobs:
 
       - name: Get commit hash
         id: commit
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Create release
         id: create_release

CMakeLists.txt

@@ -36,7 +36,6 @@ option(SD_VULKAN                     "sd: vulkan backend" OFF)
 option(SD_OPENCL                     "sd: opencl backend" OFF)
 option(SD_SYCL                       "sd: sycl backend" OFF)
 option(SD_MUSA                       "sd: musa backend" OFF)
-option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
 option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" OFF)
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
@@ -70,18 +69,12 @@ if (SD_HIPBLAS)
     message("-- Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIP ON)
     add_definitions(-DSD_USE_CUDA)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
-    endif()
 endif ()
 
 if(SD_MUSA)
     message("-- Use MUSA as backend stable-diffusion")
     set(GGML_MUSA ON)
     add_definitions(-DSD_USE_CUDA)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
-    endif()
 endif()
 
 set(SD_LIB stable-diffusion)

docs/anima.md

@@ -5,6 +5,7 @@
 - Download Anima
     - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/diffusion_models
     - gguf: https://huggingface.co/Bedovyy/Anima-GGUF/tree/main
+    - gguf Anima2: https://huggingface.co/JusteLeo/Anima2-GGUF/tree/main
 - Download vae
     - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
 - Download Qwen3-0.6B-Base
@@ -17,4 +18,4 @@
 .\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\anima-preview.safetensors --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
 ```
 
-<img alt="anima image example" src="../assets/anima/example.png" />
+<img alt="anima image example" src="../assets/anima/example.png" />

docs/caching.md

@@ -11,6 +11,7 @@ Caching methods accelerate diffusion inference by reusing intermediate computati
 | `dbcache` | DiT models | Block-level L1 residual threshold |
 | `taylorseer` | DiT models | Taylor series approximation |
 | `cache-dit` | DiT models | Combined DBCache + TaylorSeer |
+| `spectrum` | UNET models | Chebyshev + Taylor output forecasting |
 
 ### UCache (UNET Models)
 
@@ -79,7 +80,7 @@ Uses Taylor series approximation to predict block outputs:
 Combines DBCache and TaylorSeer:
 
 ```bash
---cache-mode cache-dit --cache-preset fast
+--cache-mode cache-dit
 ```
 
 #### Parameters
@@ -91,14 +92,6 @@ Combines DBCache and TaylorSeer:
 | `threshold` | L1 residual difference threshold | 0.08 |
 | `warmup` | Steps before caching starts | 8 |
 
-#### Presets
-
-Available presets: `slow`, `medium`, `fast`, `ultra` (or `s`, `m`, `f`, `u`).
-
-```bash
---cache-mode cache-dit --cache-preset fast
-```
-
 #### SCM Options
 
 Steps Computation Mask controls which steps can be cached:
@@ -118,6 +111,28 @@ Mask values: `1` = compute, `0` = can cache.
 --scm-policy dynamic
 ```
 
+### Spectrum (UNET Models)
+
+Spectrum uses Chebyshev polynomial fitting blended with Taylor extrapolation to predict denoised outputs, skipping entire UNet forward passes. Based on the paper [Spectrum: Adaptive Spectral Feature Forecasting for Efficient Diffusion Sampling](https://github.com/tingyu215/Spectrum).
+
+```bash
+sd-cli -m model.safetensors -p "a cat" --cache-mode spectrum
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `w` | Chebyshev vs Taylor blend weight (0=Taylor, 1=Chebyshev) | 0.40 |
+| `m` | Chebyshev polynomial degree | 3 |
+| `lam` | Ridge regression regularization | 1.0 |
+| `window` | Initial window size (compute every N steps) | 2 |
+| `flex` | Window growth per computed step after warmup | 0.50 |
+| `warmup` | Steps to always compute before caching starts | 4 |
+| `stop` | Stop caching at this fraction of total steps | 0.9 |
+
+```
+
 ### Performance Tips
 
 - Start with default thresholds and adjust based on output quality

examples/cli/README.md

@@ -4,11 +4,12 @@
 usage: ./bin/sd-cli  [options]
 
 CLI Options:
-  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png)
-  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
+  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default:
+                              ./output.png) (eg. output_%03d.png)
   --preview-path <string>     path to write preview image to (default: ./preview.png)
   --preview-interval <int>    interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at
                               every step)
+  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
   --canny                     apply canny preprocessor (edge detection)
   --convert-name              convert tensor name (for convert mode)
   -v, --verbose               print extra info
@@ -59,6 +60,7 @@ Context Options:
   --circularx                              enable circular RoPE wrapping on x-axis (width) only
   --circulary                              enable circular RoPE wrapping on y-axis (height) only
   --chroma-disable-dit-mask                disable dit mask for chroma
+  --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --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
@@ -107,7 +109,7 @@ Generation Options:
                                            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/TCD/res_multistep/res_2s (default: 0)
+  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --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)
@@ -115,7 +117,7 @@ Generation Options:
   --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/TCD/res_multistep/res_2s (default: 0)
+  --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
@@ -124,21 +126,24 @@ Generation Options:
   --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,
-                                           res_multistep, res_2s] (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, res_multistep, res_2s] default: euler for Flux/SD3/Wan, euler_a otherwise
+  --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, res_multistep, res_2s] (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, res_multistep, res_2s] default: euler for Flux/SD3/Wan,
+                                           euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
                                            kl_optimal, lcm, bong_tangent], 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)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level),
+                                           'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples:
-                                           "threshold=0.25" or "threshold=1.5,reset=0"
-  --cache-preset                           cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=;
+                                           spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=. Examples:
+                                           "threshold=0.25" or "threshold=1.5,reset=0" or "w=0.4,window=2"
   --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
 ```

examples/common/common.hpp

@@ -1047,7 +1047,6 @@ struct SDGenerationParams {
 
     std::string cache_mode;
     std::string cache_option;
-    std::string cache_preset;
     std::string scm_mask;
     bool scm_policy_dynamic = true;
     sd_cache_params_t cache_params{};
@@ -1422,8 +1421,8 @@ struct SDGenerationParams {
             }
             cache_mode = argv_to_utf8(index, argv);
             if (cache_mode != "easycache" && cache_mode != "ucache" &&
-                cache_mode != "dbcache" && cache_mode != "taylorseer" && cache_mode != "cache-dit") {
-                fprintf(stderr, "error: invalid cache mode '%s', must be 'easycache', 'ucache', 'dbcache', 'taylorseer', or 'cache-dit'\n", cache_mode.c_str());
+                cache_mode != "dbcache" && cache_mode != "taylorseer" && cache_mode != "cache-dit" && cache_mode != "spectrum") {
+                fprintf(stderr, "error: invalid cache mode '%s', must be 'easycache', 'ucache', 'dbcache', 'taylorseer', 'cache-dit', or 'spectrum'\n", cache_mode.c_str());
                 return -1;
             }
             return 1;
@@ -1461,21 +1460,6 @@ struct SDGenerationParams {
             return 1;
         };
 
-        auto on_cache_preset_arg = [&](int argc, const char** argv, int index) {
-            if (++index >= argc) {
-                return -1;
-            }
-            cache_preset = argv_to_utf8(index, argv);
-            if (cache_preset != "slow" && cache_preset != "s" && cache_preset != "S" &&
-                cache_preset != "medium" && cache_preset != "m" && cache_preset != "M" &&
-                cache_preset != "fast" && cache_preset != "f" && cache_preset != "F" &&
-                cache_preset != "ultra" && cache_preset != "u" && cache_preset != "U") {
-                fprintf(stderr, "error: invalid cache preset '%s', must be 'slow'/'s', 'medium'/'m', 'fast'/'f', or 'ultra'/'u'\n", cache_preset.c_str());
-                return -1;
-            }
-            return 1;
-        };
-
         options.manual_options = {
             {"-s",
              "--seed",
@@ -1513,16 +1497,12 @@ struct SDGenerationParams {
              on_ref_image_arg},
             {"",
              "--cache-mode",
-             "caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level)",
+             "caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)",
              on_cache_mode_arg},
             {"",
              "--cache-option",
-             "named cache params (key=value format, comma-separated). easycache/ucache: threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples: \"threshold=0.25\" or \"threshold=1.5,reset=0\"",
+             "named cache params (key=value format, comma-separated). easycache/ucache: threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=. Examples: \"threshold=0.25\" or \"threshold=1.5,reset=0\"",
              on_cache_option_arg},
-            {"",
-             "--cache-preset",
-             "cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'",
-             on_cache_preset_arg},
             {"",
              "--scm-mask",
              "SCM steps mask for cache-dit: comma-separated 0/1 (e.g., \"1,1,1,0,0,1,0,0,1,0\") - 1=compute, 0=can cache",
@@ -1575,7 +1555,6 @@ struct SDGenerationParams {
         load_if_exists("negative_prompt", negative_prompt);
         load_if_exists("cache_mode", cache_mode);
         load_if_exists("cache_option", cache_option);
-        load_if_exists("cache_preset", cache_preset);
         load_if_exists("scm_mask", scm_mask);
 
         load_if_exists("clip_skip", clip_skip);
@@ -1779,7 +1758,23 @@ struct SDGenerationParams {
                     } else if (key == "Bn" || key == "bn") {
                         cache_params.Bn_compute_blocks = std::stoi(val);
                     } else if (key == "warmup") {
-                        cache_params.max_warmup_steps = std::stoi(val);
+                        if (cache_mode == "spectrum") {
+                            cache_params.spectrum_warmup_steps = std::stoi(val);
+                        } else {
+                            cache_params.max_warmup_steps = std::stoi(val);
+                        }
+                    } else if (key == "w") {
+                        cache_params.spectrum_w = std::stof(val);
+                    } else if (key == "m") {
+                        cache_params.spectrum_m = std::stoi(val);
+                    } else if (key == "lam") {
+                        cache_params.spectrum_lam = std::stof(val);
+                    } else if (key == "window") {
+                        cache_params.spectrum_window_size = std::stoi(val);
+                    } else if (key == "flex") {
+                        cache_params.spectrum_flex_window = std::stof(val);
+                    } else if (key == "stop") {
+                        cache_params.spectrum_stop_percent = std::stof(val);
                     } else {
                         LOG_ERROR("error: unknown cache parameter '%s'", key.c_str());
                         return false;
@@ -1794,39 +1789,17 @@ struct SDGenerationParams {
 
         if (!cache_mode.empty()) {
             if (cache_mode == "easycache") {
-                cache_params.mode                   = SD_CACHE_EASYCACHE;
-                cache_params.reuse_threshold        = 0.2f;
-                cache_params.start_percent          = 0.15f;
-                cache_params.end_percent            = 0.95f;
-                cache_params.error_decay_rate       = 1.0f;
-                cache_params.use_relative_threshold = true;
-                cache_params.reset_error_on_compute = true;
+                cache_params.mode = SD_CACHE_EASYCACHE;
             } else if (cache_mode == "ucache") {
-                cache_params.mode                   = SD_CACHE_UCACHE;
-                cache_params.reuse_threshold        = 1.0f;
-                cache_params.start_percent          = 0.15f;
-                cache_params.end_percent            = 0.95f;
-                cache_params.error_decay_rate       = 1.0f;
-                cache_params.use_relative_threshold = true;
-                cache_params.reset_error_on_compute = true;
+                cache_params.mode = SD_CACHE_UCACHE;
             } else if (cache_mode == "dbcache") {
-                cache_params.mode                    = SD_CACHE_DBCACHE;
-                cache_params.Fn_compute_blocks       = 8;
-                cache_params.Bn_compute_blocks       = 0;
-                cache_params.residual_diff_threshold = 0.08f;
-                cache_params.max_warmup_steps        = 8;
+                cache_params.mode = SD_CACHE_DBCACHE;
             } else if (cache_mode == "taylorseer") {
-                cache_params.mode                    = SD_CACHE_TAYLORSEER;
-                cache_params.Fn_compute_blocks       = 8;
-                cache_params.Bn_compute_blocks       = 0;
-                cache_params.residual_diff_threshold = 0.08f;
-                cache_params.max_warmup_steps        = 8;
+                cache_params.mode = SD_CACHE_TAYLORSEER;
             } else if (cache_mode == "cache-dit") {
-                cache_params.mode                    = SD_CACHE_CACHE_DIT;
-                cache_params.Fn_compute_blocks       = 8;
-                cache_params.Bn_compute_blocks       = 0;
-                cache_params.residual_diff_threshold = 0.08f;
-                cache_params.max_warmup_steps        = 8;
+                cache_params.mode = SD_CACHE_CACHE_DIT;
+            } else if (cache_mode == "spectrum") {
+                cache_params.mode = SD_CACHE_SPECTRUM;
             }
 
             if (!cache_option.empty()) {

examples/server/README.md

@@ -4,12 +4,12 @@
 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)
-  --serve-html-path <string>  path to HTML file to serve at root (optional)
-  -v, --verbose               print extra info
-  --color                     colors the logging tags according to level
-  -h, --help                  show this help message and exit
+  -l, --listen-ip <string>      server listen ip (default: 127.0.0.1)        
+  --serve-html-path <string>    path to HTML file to serve at root (optional)
+  --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
@@ -39,10 +39,10 @@ Context Options:
   --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
+  --mmap                                   whether to memory-map model
   --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)
-  --mmap                                   whether to memory-map model
   --fa                                     use flash attention
   --diffusion-fa                           use flash attention in the diffusion model only
   --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
@@ -51,6 +51,7 @@ Context Options:
   --circularx                              enable circular RoPE wrapping on x-axis (width) only
   --circulary                              enable circular RoPE wrapping on y-axis (height) only
   --chroma-disable-dit-mask                disable dit mask for chroma
+  --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --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
@@ -99,7 +100,7 @@ Default Generation Options:
                                            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/TCD/res_multistep/res_2s (default: 0)
+  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --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)
@@ -107,7 +108,7 @@ Default Generation Options:
   --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/TCD/res_multistep/res_2s (default: 0)
+  --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
@@ -116,21 +117,22 @@ Default Generation Options:
   --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,
-                                           res_multistep, res_2s] (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, res_multistep, res_2s] default: euler for Flux/SD3/Wan, euler_a otherwise
+  --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, res_multistep, res_2s] (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, res_multistep, res_2s] default: euler for Flux/SD3/Wan,
+                                           euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
                                            kl_optimal, lcm, bong_tangent], 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)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
                                            threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples:
                                            "threshold=0.25" or "threshold=1.5,reset=0"
-  --cache-preset                           cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'
   --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
 ```

examples/server/main.cpp

@@ -266,8 +266,21 @@ void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
 struct LoraEntry {
     std::string name;
     std::string path;
+    std::string fullpath;
 };
 
+void free_results(sd_image_t* result_images, int num_results) {
+    if (result_images) {
+        for (int i = 0; i < num_results; ++i) {
+            if (result_images[i].data) {
+                stbi_image_free(result_images[i].data);
+                result_images[i].data = nullptr;
+            }
+        }
+    }
+    free(result_images);
+}
+
 int main(int argc, const char** argv) {
     if (argc > 1 && std::string(argv[1]) == "--version") {
         std::cout << version_string() << "\n";
@@ -321,7 +334,8 @@ int main(int argc, const char** argv) {
 
                 LoraEntry e;
                 e.name          = p.stem().u8string();
-                std::string rel = fs::relative(p, lora_dir).u8string();
+                e.fullpath      = p.u8string();
+                std::string rel = p.lexically_relative(lora_dir).u8string();
                 std::replace(rel.begin(), rel.end(), '\\', '/');
                 e.path = rel;
 
@@ -340,10 +354,11 @@ int main(int argc, const char** argv) {
         }
     };
 
-    auto is_valid_lora_path = [&](const std::string& path) -> bool {
+    auto get_lora_full_path = [&](const std::string& path) -> std::string {
         std::lock_guard<std::mutex> lock(lora_mutex);
-        return std::any_of(lora_cache.begin(), lora_cache.end(),
-                           [&](const LoraEntry& e) { return e.path == path; });
+        auto it = std::find_if(lora_cache.begin(), lora_cache.end(),
+                               [&](const LoraEntry& e) { return e.path == path; });
+        return (it != lora_cache.end()) ? it->fullpath : "";
     };
 
     httplib::Server svr;
@@ -534,6 +549,7 @@ int main(int argc, const char** argv) {
                 item["b64_json"] = b64;
                 out["data"].push_back(item);
             }
+            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;
@@ -564,8 +580,9 @@ int main(int argc, const char** argv) {
 
             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) {
+            size_t image_count    = req.form.get_file_count("image[]");
+            bool has_legacy_image = req.form.has_file("image");
+            if (image_count == 0 && !has_legacy_image) {
                 res.status = 400;
                 res.set_content(R"({"error":"at least one image[] required"})", "application/json");
                 return;
@@ -576,6 +593,10 @@ int main(int argc, const char** argv) {
                 auto file = req.form.get_file("image[]", i);
                 images_bytes.emplace_back(file.content.begin(), file.content.end());
             }
+            if (image_count == 0 && has_legacy_image) {
+                auto file = req.form.get_file("image");
+                images_bytes.emplace_back(file.content.begin(), file.content.end());
+            }
 
             std::vector<uint8_t> mask_bytes;
             if (req.form.has_file("mask")) {
@@ -773,6 +794,7 @@ int main(int argc, const char** argv) {
                 item["b64_json"] = b64;
                 out["data"].push_back(item);
             }
+            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;
@@ -862,11 +884,12 @@ int main(int argc, const char** argv) {
                         return bad("lora.path required");
                     }
 
-                    if (!is_valid_lora_path(path)) {
+                    std::string fullpath = get_lora_full_path(path);
+                    if (fullpath.empty()) {
                         return bad("invalid lora path: " + path);
                     }
 
-                    lora_path_storage.push_back(path);
+                    lora_path_storage.push_back(fullpath);
                     sd_lora_t l;
                     l.is_high_noise = is_high_noise;
                     l.multiplier    = multiplier;
@@ -1092,6 +1115,7 @@ int main(int argc, const char** argv) {
                 std::string b64 = base64_encode(image_bytes);
                 out["images"].push_back(b64);
             }
+            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;

include/stable-diffusion.h

@@ -251,6 +251,7 @@ enum sd_cache_mode_t {
     SD_CACHE_DBCACHE,
     SD_CACHE_TAYLORSEER,
     SD_CACHE_CACHE_DIT,
+    SD_CACHE_SPECTRUM,
 };
 
 typedef struct {
@@ -271,6 +272,13 @@ typedef struct {
     int taylorseer_skip_interval;
     const char* scm_mask;
     bool scm_policy_dynamic;
+    float spectrum_w;
+    int spectrum_m;
+    float spectrum_lam;
+    int spectrum_window_size;
+    float spectrum_flex_window;
+    int spectrum_warmup_steps;
+    float spectrum_stop_percent;
 } sd_cache_params_t;
 
 typedef struct {

src/cache_dit.hpp

@@ -603,87 +603,6 @@ inline std::vector<int> generate_scm_mask(
     return mask;
 }
 
-inline std::vector<int> get_scm_preset(const std::string& preset, int total_steps) {
-    struct Preset {
-        std::vector<int> compute_bins;
-        std::vector<int> cache_bins;
-    };
-
-    Preset slow   = {{8, 3, 3, 2, 1, 1}, {1, 2, 2, 2, 3}};
-    Preset medium = {{6, 2, 2, 2, 2, 1}, {1, 3, 3, 3, 3}};
-    Preset fast   = {{6, 1, 1, 1, 1, 1}, {1, 3, 4, 5, 4}};
-    Preset ultra  = {{4, 1, 1, 1, 1}, {2, 5, 6, 7}};
-
-    Preset* p = nullptr;
-    if (preset == "slow" || preset == "s" || preset == "S")
-        p = &slow;
-    else if (preset == "medium" || preset == "m" || preset == "M")
-        p = &medium;
-    else if (preset == "fast" || preset == "f" || preset == "F")
-        p = &fast;
-    else if (preset == "ultra" || preset == "u" || preset == "U")
-        p = &ultra;
-    else
-        return {};
-
-    if (total_steps != 28 && total_steps > 0) {
-        float scale = static_cast<float>(total_steps) / 28.0f;
-        std::vector<int> scaled_compute, scaled_cache;
-
-        for (int v : p->compute_bins) {
-            scaled_compute.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
-        }
-        for (int v : p->cache_bins) {
-            scaled_cache.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
-        }
-
-        return generate_scm_mask(scaled_compute, scaled_cache, total_steps);
-    }
-
-    return generate_scm_mask(p->compute_bins, p->cache_bins, total_steps);
-}
-
-inline float get_preset_threshold(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 0.20f;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 0.25f;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 0.30f;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 0.34f;
-    return 0.08f;
-}
-
-inline int get_preset_warmup(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 8;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 6;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 6;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 4;
-    return 8;
-}
-
-inline int get_preset_Fn(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 8;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 8;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 6;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 4;
-    return 8;
-}
-
-inline int get_preset_Bn(const std::string& preset) {
-    (void)preset;
-    return 0;
-}
-
 inline void parse_dbcache_options(const std::string& opts, DBCacheConfig& cfg) {
     if (opts.empty())
         return;

src/flux.hpp

@@ -103,11 +103,13 @@ namespace Flux {
             auto norm     = std::dynamic_pointer_cast<QKNorm>(blocks["norm"]);
 
             auto qkv         = qkv_proj->forward(ctx, x);
-            auto qkv_vec     = ggml_ext_chunk(ctx->ggml_ctx, qkv, 3, 0, true);
-            int64_t head_dim = qkv_vec[0]->ne[0] / num_heads;
-            auto q           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[0], head_dim, num_heads, qkv_vec[0]->ne[1], qkv_vec[0]->ne[2]);
-            auto k           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[1], head_dim, num_heads, qkv_vec[1]->ne[1], qkv_vec[1]->ne[2]);
-            auto v           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[2], head_dim, num_heads, qkv_vec[2]->ne[1], qkv_vec[2]->ne[2]);
+            int64_t head_dim = qkv->ne[0] / 3 / num_heads;
+            auto q           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], 0);
+            auto k           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], (qkv->nb[0]) * qkv->ne[0] / 3);
+            auto v           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], (qkv->nb[0]) * 2 * qkv->ne[0] / 3);
             q                = norm->query_norm(ctx, q);
             k                = norm->key_norm(ctx, k);
             return {q, k, v};
@@ -491,15 +493,14 @@ namespace Flux {
             auto x_mod   = Flux::modulate(ctx->ggml_ctx, pre_norm->forward(ctx, x), mod.shift, mod.scale);
             auto qkv_mlp = linear1->forward(ctx, x_mod);  // [N, n_token, hidden_size * 3 + mlp_hidden_dim*mlp_mult_factor]
 
-            auto q = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], 0);
-            auto k = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], hidden_size * qkv_mlp->nb[0]);
-            auto v = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], hidden_size * 2 * qkv_mlp->nb[0]);
-
             int64_t head_dim = hidden_size / num_heads;
 
-            q = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, q), head_dim, num_heads, q->ne[1], q->ne[2]);  // [N, n_token, n_head, d_head]
-            k = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, k), head_dim, num_heads, k->ne[1], k->ne[2]);  // [N, n_token, n_head, d_head]
-            v = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, v), head_dim, num_heads, v->ne[1], v->ne[2]);  // [N, n_token, n_head, d_head]
+            auto q = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], 0);
+            auto k = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], (qkv_mlp->nb[0]) * hidden_size);
+            auto v = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], (qkv_mlp->nb[0]) * 2 * hidden_size);
 
             q         = norm->query_norm(ctx, q);
             k         = norm->key_norm(ctx, k);

src/ggml_extend.hpp

@@ -491,12 +491,16 @@ __STATIC_INLINE__ void ggml_ext_tensor_split_2d(struct ggml_tensor* input,
     int64_t height   = output->ne[1];
     int64_t channels = output->ne[2];
     int64_t ne3      = output->ne[3];
+
+    int64_t input_width  = input->ne[0];
+    int64_t input_height = input->ne[1];
+
     GGML_ASSERT(input->type == GGML_TYPE_F32 && output->type == GGML_TYPE_F32);
     for (int iy = 0; iy < height; iy++) {
         for (int ix = 0; ix < width; ix++) {
             for (int k = 0; k < channels; k++) {
                 for (int l = 0; l < ne3; l++) {
-                    float value = ggml_ext_tensor_get_f32(input, ix + x, iy + y, k, l);
+                    float value = ggml_ext_tensor_get_f32(input, (ix + x) % input_width, (iy + y) % input_height, k, l);
                     ggml_ext_tensor_set_f32(output, value, ix, iy, k, l);
                 }
             }
@@ -516,6 +520,8 @@ __STATIC_INLINE__ void ggml_ext_tensor_merge_2d(struct ggml_tensor* input,
                                                 int y,
                                                 int overlap_x,
                                                 int overlap_y,
+                                                bool circular_x,
+                                                bool circular_y,
                                                 int x_skip = 0,
                                                 int y_skip = 0) {
     int64_t width    = input->ne[0];
@@ -533,12 +539,12 @@ __STATIC_INLINE__ void ggml_ext_tensor_merge_2d(struct ggml_tensor* input,
                 for (int l = 0; l < ne3; l++) {
                     float new_value = ggml_ext_tensor_get_f32(input, ix, iy, k, l);
                     if (overlap_x > 0 || overlap_y > 0) {  // blend colors in overlapped area
-                        float old_value = ggml_ext_tensor_get_f32(output, x + ix, y + iy, k, l);
+                        float old_value = ggml_ext_tensor_get_f32(output, (x + ix) % img_width, (y + iy) % img_height, k, l);
 
-                        const float x_f_0 = (overlap_x > 0 && x > 0) ? (ix - x_skip) / float(overlap_x) : 1;
-                        const float x_f_1 = (overlap_x > 0 && x < (img_width - width)) ? (width - ix) / float(overlap_x) : 1;
-                        const float y_f_0 = (overlap_y > 0 && y > 0) ? (iy - y_skip) / float(overlap_y) : 1;
-                        const float y_f_1 = (overlap_y > 0 && y < (img_height - height)) ? (height - iy) / float(overlap_y) : 1;
+                        const float x_f_0 = (circular_x || (overlap_x > 0 && x > 0)) ? (ix - x_skip) / float(overlap_x) : 1;
+                        const float x_f_1 = (circular_x || (overlap_x > 0 && x < (img_width - width))) ? (width - ix) / float(overlap_x) : 1;
+                        const float y_f_0 = (circular_y || (overlap_y > 0 && y > 0)) ? (iy - y_skip) / float(overlap_y) : 1;
+                        const float y_f_1 = (circular_y || (overlap_y > 0 && y < (img_height - height))) ? (height - iy) / float(overlap_y) : 1;
 
                         const float x_f = std::min(std::min(x_f_0, x_f_1), 1.f);
                         const float y_f = std::min(std::min(y_f_0, y_f_1), 1.f);
@@ -546,9 +552,9 @@ __STATIC_INLINE__ void ggml_ext_tensor_merge_2d(struct ggml_tensor* input,
                         ggml_ext_tensor_set_f32(
                             output,
                             old_value + new_value * smootherstep_f32(y_f) * smootherstep_f32(x_f),
-                            x + ix, y + iy, k, l);
+                            (x + ix) % img_width, (y + iy) % img_height, k, l);
                     } else {
-                        ggml_ext_tensor_set_f32(output, new_value, x + ix, y + iy, k, l);
+                        ggml_ext_tensor_set_f32(output, new_value, (x + ix) % img_width, (y + iy) % img_height, k, l);
                     }
                 }
             }
@@ -773,10 +779,31 @@ __STATIC_INLINE__ void sd_tiling_calc_tiles(int& num_tiles_dim,
                                             float& tile_overlap_factor_dim,
                                             int small_dim,
                                             int tile_size,
-                                            const float tile_overlap_factor) {
+                                            const float tile_overlap_factor,
+                                            bool circular) {
     int tile_overlap     = static_cast<int>(tile_size * tile_overlap_factor);
     int non_tile_overlap = tile_size - tile_overlap;
 
+    if (circular) {
+        // circular means the last and first tile are overlapping (wraping around)
+        num_tiles_dim = small_dim / non_tile_overlap;
+
+        if (num_tiles_dim < 1) {
+            num_tiles_dim = 1;
+        }
+
+        tile_overlap_factor_dim = (tile_size - small_dim / num_tiles_dim) / (float)tile_size;
+
+        // if single tile and tile_overlap_factor is not 0, add one to ensure we have at least two overlapping tiles
+        if (num_tiles_dim == 1 && tile_overlap_factor_dim > 0) {
+            num_tiles_dim++;
+            tile_overlap_factor_dim = 0.5;
+        }
+
+        return;
+    }
+    // else, non-circular means the last and first tile are not overlapping
+
     num_tiles_dim     = (small_dim - tile_overlap) / non_tile_overlap;
     int overshoot_dim = ((num_tiles_dim + 1) * non_tile_overlap + tile_overlap) % small_dim;
 
@@ -805,6 +832,8 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
                                             const int p_tile_size_x,
                                             const int p_tile_size_y,
                                             const float tile_overlap_factor,
+                                            const bool circular_x,
+                                            const bool circular_y,
                                             on_tile_process on_processing) {
     output = ggml_set_f32(output, 0);
 
@@ -829,11 +858,11 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
 
     int num_tiles_x;
     float tile_overlap_factor_x;
-    sd_tiling_calc_tiles(num_tiles_x, tile_overlap_factor_x, small_width, p_tile_size_x, tile_overlap_factor);
+    sd_tiling_calc_tiles(num_tiles_x, tile_overlap_factor_x, small_width, p_tile_size_x, tile_overlap_factor, circular_x);
 
     int num_tiles_y;
     float tile_overlap_factor_y;
-    sd_tiling_calc_tiles(num_tiles_y, tile_overlap_factor_y, small_height, p_tile_size_y, tile_overlap_factor);
+    sd_tiling_calc_tiles(num_tiles_y, tile_overlap_factor_y, small_height, p_tile_size_y, tile_overlap_factor, circular_y);
 
     LOG_DEBUG("num tiles : %d, %d ", num_tiles_x, num_tiles_y);
     LOG_DEBUG("optimal overlap : %f, %f (targeting %f)", tile_overlap_factor_x, tile_overlap_factor_y, tile_overlap_factor);
@@ -887,7 +916,7 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
     float last_time = 0.0f;
     for (int y = 0; y < small_height && !last_y; y += non_tile_overlap_y) {
         int dy = 0;
-        if (y + tile_size_y >= small_height) {
+        if (!circular_y && y + tile_size_y >= small_height) {
             int _y = y;
             y      = small_height - tile_size_y;
             dy     = _y - y;
@@ -898,7 +927,7 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
         }
         for (int x = 0; x < small_width && !last_x; x += non_tile_overlap_x) {
             int dx = 0;
-            if (x + tile_size_x >= small_width) {
+            if (!circular_x && x + tile_size_x >= small_width) {
                 int _x = x;
                 x      = small_width - tile_size_x;
                 dx     = _x - x;
@@ -919,7 +948,7 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
             int64_t t1 = ggml_time_ms();
             ggml_ext_tensor_split_2d(input, input_tile, x_in, y_in);
             if (on_processing(input_tile, output_tile, false)) {
-                ggml_ext_tensor_merge_2d(output_tile, output, x_out, y_out, overlap_x_out, overlap_y_out, dx, dy);
+                ggml_ext_tensor_merge_2d(output_tile, output, x_out, y_out, overlap_x_out, overlap_y_out, circular_x, circular_y, dx, dy);
 
                 int64_t t2 = ggml_time_ms();
                 last_time  = (t2 - t1) / 1000.0f;
@@ -942,8 +971,10 @@ __STATIC_INLINE__ void sd_tiling(ggml_tensor* input,
                                  const int scale,
                                  const int tile_size,
                                  const float tile_overlap_factor,
+                                 const bool circular_x,
+                                 const bool circular_y,
                                  on_tile_process on_processing) {
-    sd_tiling_non_square(input, output, scale, tile_size, tile_size, tile_overlap_factor, on_processing);
+    sd_tiling_non_square(input, output, scale, tile_size, tile_size, tile_overlap_factor, circular_x, circular_y, on_processing);
 }
 
 __STATIC_INLINE__ struct ggml_tensor* ggml_ext_group_norm_32(struct ggml_context* ctx,
@@ -1219,6 +1250,11 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_zeros(struct ggml_context* ctx,
     return ggml_ext_full(ctx, 0.f, ne0, ne1, ne2, ne3);
 }
 
+__STATIC_INLINE__ struct ggml_tensor* ggml_ext_zeros_like(struct ggml_context* ctx,
+                                                          struct ggml_tensor* x) {
+    return ggml_ext_zeros(ctx, x->ne[0], x->ne[1], x->ne[2], x->ne[3]);
+}
+
 __STATIC_INLINE__ struct ggml_tensor* ggml_ext_ones(struct ggml_context* ctx,
                                                     int64_t ne0,
                                                     int64_t ne1,
@@ -1227,6 +1263,11 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_ones(struct ggml_context* ctx,
     return ggml_ext_full(ctx, 1.f, ne0, ne1, ne2, ne3);
 }
 
+__STATIC_INLINE__ struct ggml_tensor* ggml_ext_ones_like(struct ggml_context* ctx,
+                                                         struct ggml_tensor* x) {
+    return ggml_ext_ones(ctx, x->ne[0], x->ne[1], x->ne[2], x->ne[3]);
+}
+
 __STATIC_INLINE__ ggml_tensor* ggml_ext_cast_f32(ggml_context* ctx, ggml_tensor* a) {
 #ifdef SD_USE_VULKAN
     auto zero_index = ggml_get_tensor(ctx, "ggml_runner_build_in_tensor:zero_int");

src/model.cpp

@@ -1057,7 +1057,7 @@ SDVersion ModelLoader::get_sd_version() {
             if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
                 return VERSION_QWEN_IMAGE;
             }
-            if (tensor_storage.name.find("model.diffusion_model.net.llm_adapter.blocks.0.cross_attn.q_proj.weight") != std::string::npos) {
+            if (tensor_storage.name.find("llm_adapter.blocks.0.cross_attn.q_proj.weight") != std::string::npos) {
                 return VERSION_ANIMA;
             }
             if (tensor_storage.name.find("model.diffusion_model.double_stream_modulation_img.lin.weight") != std::string::npos) {

src/name_conversion.cpp

@@ -653,6 +653,14 @@ std::string convert_diffusers_dit_to_original_lumina2(std::string name) {
     return name;
 }
 
+std::string convert_other_dit_to_original_anima(std::string name) {
+    static const std::string anima_net_prefix = "net.";
+    if (!starts_with(name, anima_net_prefix)) {
+        name = anima_net_prefix + name;
+    }
+    return name;
+}
+
 std::string convert_diffusion_model_name(std::string name, std::string prefix, SDVersion version) {
     if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
         name = convert_diffusers_unet_to_original_sd1(name);
@@ -664,6 +672,8 @@ std::string convert_diffusion_model_name(std::string name, std::string prefix, S
         name = convert_diffusers_dit_to_original_flux(name);
     } else if (sd_version_is_z_image(version)) {
         name = convert_diffusers_dit_to_original_lumina2(name);
+    } else if (sd_version_is_anima(version)) {
+        name = convert_other_dit_to_original_anima(name);
     }
     return name;
 }
@@ -1094,14 +1104,6 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
         }
     }
 
-    if (is_lora && sd_version_is_anima(version)) {
-        static const std::string anima_diffusion_prefix = "model.diffusion_model.";
-        static const std::string anima_net_prefix       = "model.diffusion_model.net.";
-        if (starts_with(name, anima_diffusion_prefix) && !starts_with(name, anima_net_prefix)) {
-            name = anima_net_prefix + name.substr(anima_diffusion_prefix.size());
-        }
-    }
-
     // cond_stage_model
     {
         for (const auto& prefix : cond_stage_model_prefix_vec) {

src/qwen_image.hpp

@@ -404,7 +404,7 @@ namespace Qwen {
 
             auto t_emb = time_text_embed->forward(ctx, timestep);
             if (params.zero_cond_t) {
-                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros(ctx->ggml_ctx, timestep->ne[0], timestep->ne[1], timestep->ne[2], timestep->ne[3]));
+                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep));
                 t_emb        = ggml_concat(ctx->ggml_ctx, t_emb, t_emb_0, 1);
             }
             auto img = img_in->forward(ctx, x);

src/spectrum.hpp

@@ -0,0 +1,195 @@
+#ifndef __SPECTRUM_HPP__
+#define __SPECTRUM_HPP__
+
+#include <cmath>
+#include <cstring>
+#include <vector>
+
+#include "ggml_extend.hpp"
+
+struct SpectrumConfig {
+    float w            = 0.40f;
+    int m              = 3;
+    float lam          = 1.0f;
+    int window_size    = 2;
+    float flex_window  = 0.50f;
+    int warmup_steps   = 4;
+    float stop_percent = 0.9f;
+};
+
+struct SpectrumState {
+    SpectrumConfig config;
+    int cnt                 = 0;
+    int num_cached          = 0;
+    float curr_ws           = 2.0f;
+    int K                   = 6;
+    int stop_step           = 0;
+    int total_steps_skipped = 0;
+
+    std::vector<std::vector<float>> H_buf;
+    std::vector<float> T_buf;
+
+    void init(const SpectrumConfig& cfg, size_t total_steps) {
+        config              = cfg;
+        cnt                 = 0;
+        num_cached          = 0;
+        curr_ws             = (float)cfg.window_size;
+        K                   = std::max(cfg.m + 1, 6);
+        stop_step           = (int)(cfg.stop_percent * (float)total_steps);
+        total_steps_skipped = 0;
+        H_buf.clear();
+        T_buf.clear();
+    }
+
+    float taus(int step_cnt) const {
+        return (step_cnt / 50.0f) * 2.0f - 1.0f;
+    }
+
+    bool should_predict() {
+        if (cnt < config.warmup_steps)
+            return false;
+        if (stop_step > 0 && cnt >= stop_step)
+            return false;
+        if ((int)H_buf.size() < 2)
+            return false;
+
+        int ws = std::max(1, (int)std::floor(curr_ws));
+        return (num_cached + 1) % ws != 0;
+    }
+
+    void update(const struct ggml_tensor* denoised) {
+        int64_t ne        = ggml_nelements(denoised);
+        const float* data = (const float*)denoised->data;
+
+        H_buf.emplace_back(data, data + ne);
+        T_buf.push_back(taus(cnt));
+
+        while ((int)H_buf.size() > K) {
+            H_buf.erase(H_buf.begin());
+            T_buf.erase(T_buf.begin());
+        }
+
+        if (cnt >= config.warmup_steps)
+            curr_ws += config.flex_window;
+
+        num_cached = 0;
+        cnt++;
+    }
+
+    void predict(struct ggml_tensor* denoised) {
+        int64_t F    = (int64_t)H_buf[0].size();
+        int K_curr   = (int)H_buf.size();
+        int M1       = config.m + 1;
+        float tau_at = taus(cnt);
+
+        // Design matrix X: K_curr x M1 (Chebyshev basis)
+        std::vector<float> X(K_curr * M1);
+        for (int i = 0; i < K_curr; i++) {
+            X[i * M1] = 1.0f;
+            if (M1 > 1)
+                X[i * M1 + 1] = T_buf[i];
+            for (int j = 2; j < M1; j++)
+                X[i * M1 + j] = 2.0f * T_buf[i] * X[i * M1 + j - 1] - X[i * M1 + j - 2];
+        }
+
+        // x_star: Chebyshev basis at current tau
+        std::vector<float> x_star(M1);
+        x_star[0] = 1.0f;
+        if (M1 > 1)
+            x_star[1] = tau_at;
+        for (int j = 2; j < M1; j++)
+            x_star[j] = 2.0f * tau_at * x_star[j - 1] - x_star[j - 2];
+
+        // XtX = X^T X + lambda I
+        std::vector<float> XtX(M1 * M1, 0.0f);
+        for (int i = 0; i < M1; i++) {
+            for (int j = 0; j < M1; j++) {
+                float sum = 0.0f;
+                for (int k = 0; k < K_curr; k++)
+                    sum += X[k * M1 + i] * X[k * M1 + j];
+                XtX[i * M1 + j] = sum + (i == j ? config.lam : 0.0f);
+            }
+        }
+
+        // Cholesky decomposition
+        std::vector<float> L(M1 * M1, 0.0f);
+        if (!cholesky_decompose(XtX.data(), L.data(), M1)) {
+            float trace = 0.0f;
+            for (int i = 0; i < M1; i++)
+                trace += XtX[i * M1 + i];
+            for (int i = 0; i < M1; i++)
+                XtX[i * M1 + i] += 1e-4f * trace / M1;
+            cholesky_decompose(XtX.data(), L.data(), M1);
+        }
+
+        // Solve XtX v = x_star
+        std::vector<float> v(M1);
+        cholesky_solve(L.data(), x_star.data(), v.data(), M1);
+
+        // Prediction weights per history entry
+        std::vector<float> weights(K_curr, 0.0f);
+        for (int k = 0; k < K_curr; k++)
+            for (int j = 0; j < M1; j++)
+                weights[k] += X[k * M1 + j] * v[j];
+
+        // Blend Chebyshev and Taylor predictions
+        float* out          = (float*)denoised->data;
+        float w_cheb        = config.w;
+        float w_taylor      = 1.0f - w_cheb;
+        const float* h_last = H_buf.back().data();
+        const float* h_prev = H_buf[H_buf.size() - 2].data();
+
+        for (int64_t f = 0; f < F; f++) {
+            float pred_cheb = 0.0f;
+            for (int k = 0; k < K_curr; k++)
+                pred_cheb += weights[k] * H_buf[k][f];
+
+            float pred_taylor = h_last[f] + 0.5f * (h_last[f] - h_prev[f]);
+
+            out[f] = w_taylor * pred_taylor + w_cheb * pred_cheb;
+        }
+
+        num_cached++;
+        total_steps_skipped++;
+        cnt++;
+    }
+
+private:
+    static bool cholesky_decompose(const float* A, float* L, int n) {
+        std::memset(L, 0, n * n * sizeof(float));
+        for (int i = 0; i < n; i++) {
+            for (int j = 0; j <= i; j++) {
+                float sum = 0.0f;
+                for (int k = 0; k < j; k++)
+                    sum += L[i * n + k] * L[j * n + k];
+                if (i == j) {
+                    float diag = A[i * n + i] - sum;
+                    if (diag <= 0.0f)
+                        return false;
+                    L[i * n + j] = std::sqrt(diag);
+                } else {
+                    L[i * n + j] = (A[i * n + j] - sum) / L[j * n + j];
+                }
+            }
+        }
+        return true;
+    }
+
+    static void cholesky_solve(const float* L, const float* b, float* x, int n) {
+        std::vector<float> y(n);
+        for (int i = 0; i < n; i++) {
+            float sum = 0.0f;
+            for (int j = 0; j < i; j++)
+                sum += L[i * n + j] * y[j];
+            y[i] = (b[i] - sum) / L[i * n + i];
+        }
+        for (int i = n - 1; i >= 0; i--) {
+            float sum = 0.0f;
+            for (int j = i + 1; j < n; j++)
+                sum += L[j * n + i] * x[j];
+            x[i] = (y[i] - sum) / L[i * n + i];
+        }
+    }
+};
+
+#endif  // __SPECTRUM_HPP__

src/stable-diffusion.cpp

@@ -16,6 +16,7 @@
 #include "esrgan.hpp"
 #include "lora.hpp"
 #include "pmid.hpp"
+#include "spectrum.hpp"
 #include "tae.hpp"
 #include "ucache.hpp"
 #include "vae.hpp"
@@ -97,6 +98,19 @@ void suppress_pp(int step, int steps, float time, void* data) {
     return;
 }
 
+static float get_cache_reuse_threshold(const sd_cache_params_t& params) {
+    float reuse_threshold = params.reuse_threshold;
+    if (reuse_threshold == INFINITY) {
+        if (params.mode == SD_CACHE_EASYCACHE) {
+            reuse_threshold = 0.2;
+        }
+        else if (params.mode == SD_CACHE_UCACHE) {
+            reuse_threshold = 1.0;
+        }
+    }
+    return std::max(0.0f, reuse_threshold);
+}
+
 /*=============================================== StableDiffusionGGML ================================================*/
 
 class StableDiffusionGGML {
@@ -111,6 +125,9 @@ public:
     bool external_vae_is_invalid = false;
     bool free_params_immediately = false;
 
+    bool circular_x = false;
+    bool circular_y = false;
+
     std::shared_ptr<RNG> rng         = std::make_shared<PhiloxRNG>();
     std::shared_ptr<RNG> sampler_rng = nullptr;
     int n_threads                    = -1;
@@ -759,12 +776,8 @@ public:
             if (control_net) {
                 control_net->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
             }
-            if (first_stage_model) {
-                first_stage_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
-            }
-            if (tae_first_stage) {
-                tae_first_stage->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
-            }
+            circular_x = sd_ctx_params->circular_x;
+            circular_y = sd_ctx_params->circular_y;
         }
 
         struct ggml_init_params params;
@@ -1098,6 +1111,18 @@ public:
         cond_stage_lora_models.clear();
         diffusion_lora_models.clear();
         first_stage_lora_models.clear();
+        if (cond_stage_model) {
+            cond_stage_model->set_weight_adapter(nullptr);
+        }
+        if (diffusion_model) {
+            diffusion_model->set_weight_adapter(nullptr);
+        }
+        if (high_noise_diffusion_model) {
+            high_noise_diffusion_model->set_weight_adapter(nullptr);
+        }
+        if (first_stage_model) {
+            first_stage_model->set_weight_adapter(nullptr);
+        }
         if (lora_state.empty()) {
             return;
         }
@@ -1467,7 +1492,7 @@ public:
         sd_progress_cb_t cb = sd_get_progress_callback();
         void* cbd           = sd_get_progress_callback_data();
         sd_set_progress_callback((sd_progress_cb_t)suppress_pp, nullptr);
-        sd_tiling(input, output, scale, tile_size, tile_overlap_factor, on_processing);
+        sd_tiling(input, output, scale, tile_size, tile_overlap_factor, circular_x, circular_y, on_processing);
         sd_set_progress_callback(cb, cbd);
     }
 
@@ -1676,9 +1701,11 @@ public:
         EasyCacheState easycache_state;
         UCacheState ucache_state;
         CacheDitConditionState cachedit_state;
+        SpectrumState spectrum_state;
         bool easycache_enabled = false;
         bool ucache_enabled    = false;
         bool cachedit_enabled  = false;
+        bool spectrum_enabled  = false;
 
         if (cache_params != nullptr && cache_params->mode != SD_CACHE_DISABLED) {
             bool percent_valid = true;
@@ -1701,7 +1728,7 @@ public:
                 } else {
                     EasyCacheConfig easycache_config;
                     easycache_config.enabled         = true;
-                    easycache_config.reuse_threshold = std::max(0.0f, cache_params->reuse_threshold);
+                    easycache_config.reuse_threshold = get_cache_reuse_threshold(*cache_params);
                     easycache_config.start_percent   = cache_params->start_percent;
                     easycache_config.end_percent     = cache_params->end_percent;
                     easycache_state.init(easycache_config, denoiser.get());
@@ -1722,7 +1749,7 @@ public:
                 } else {
                     UCacheConfig ucache_config;
                     ucache_config.enabled                = true;
-                    ucache_config.reuse_threshold        = std::max(0.0f, cache_params->reuse_threshold);
+                    ucache_config.reuse_threshold        = get_cache_reuse_threshold(*cache_params);
                     ucache_config.start_percent          = cache_params->start_percent;
                     ucache_config.end_percent            = cache_params->end_percent;
                     ucache_config.error_decay_rate       = std::max(0.0f, std::min(1.0f, cache_params->error_decay_rate));
@@ -1782,6 +1809,27 @@ public:
                         LOG_WARN("CacheDIT requested but could not be initialized for this run");
                     }
                 }
+            } else if (cache_params->mode == SD_CACHE_SPECTRUM) {
+                bool spectrum_supported = sd_version_is_unet(version) || sd_version_is_dit(version);
+                if (!spectrum_supported) {
+                    LOG_WARN("Spectrum requested but not supported for this model type (only UNET and DiT models)");
+                } else {
+                    SpectrumConfig spectrum_config;
+                    spectrum_config.w            = cache_params->spectrum_w;
+                    spectrum_config.m            = cache_params->spectrum_m;
+                    spectrum_config.lam          = cache_params->spectrum_lam;
+                    spectrum_config.window_size  = cache_params->spectrum_window_size;
+                    spectrum_config.flex_window  = cache_params->spectrum_flex_window;
+                    spectrum_config.warmup_steps = cache_params->spectrum_warmup_steps;
+                    spectrum_config.stop_percent = cache_params->spectrum_stop_percent;
+                    size_t total_steps           = sigmas.size() > 0 ? sigmas.size() - 1 : 0;
+                    spectrum_state.init(spectrum_config, total_steps);
+                    spectrum_enabled = true;
+                    LOG_INFO("Spectrum enabled - w: %.2f, m: %d, lam: %.2f, window: %d, flex: %.2f, warmup: %d, stop: %.0f%%",
+                             spectrum_config.w, spectrum_config.m, spectrum_config.lam,
+                             spectrum_config.window_size, spectrum_config.flex_window,
+                             spectrum_config.warmup_steps, spectrum_config.stop_percent * 100.0f);
+                }
             }
         }
 
@@ -2004,7 +2052,29 @@ public:
                 timesteps_vec.assign(1, t);
             }
 
-            timesteps_vec  = process_timesteps(timesteps_vec, init_latent, denoise_mask);
+            timesteps_vec = process_timesteps(timesteps_vec, init_latent, denoise_mask);
+
+            if (spectrum_enabled && spectrum_state.should_predict()) {
+                spectrum_state.predict(denoised);
+
+                if (denoise_mask != nullptr) {
+                    apply_mask(denoised, init_latent, denoise_mask);
+                }
+
+                if (sd_preview_cb != nullptr && sd_should_preview_denoised()) {
+                    if (step % sd_get_preview_interval() == 0) {
+                        preview_image(work_ctx, step, denoised, version, sd_preview_mode, preview_tensor, sd_preview_cb, sd_preview_cb_data, false);
+                    }
+                }
+
+                int64_t t1 = ggml_time_us();
+                if (step > 0 || step == -(int)steps) {
+                    int showstep = std::abs(step);
+                    pretty_progress(showstep, (int)steps, (t1 - t0) / 1000000.f / showstep);
+                }
+                return denoised;
+            }
+
             auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
             std::vector<float> guidance_vec(1, guidance.distilled_guidance);
             auto guidance_tensor = vector_to_ggml_tensor(work_ctx, guidance_vec);
@@ -2178,6 +2248,10 @@ public:
                 vec_denoised[i] = latent_result * c_out + vec_input[i] * c_skip;
             }
 
+            if (spectrum_enabled) {
+                spectrum_state.update(denoised);
+            }
+
             if (denoise_mask != nullptr) {
                 apply_mask(denoised, init_latent, denoise_mask);
             }
@@ -2269,6 +2343,14 @@ public:
             }
         }
 
+        if (spectrum_enabled && spectrum_state.total_steps_skipped > 0) {
+            size_t total_steps = sigmas.size() > 0 ? sigmas.size() - 1 : 0;
+            double speedup     = static_cast<double>(total_steps) /
+                             static_cast<double>(total_steps - spectrum_state.total_steps_skipped);
+            LOG_INFO("Spectrum skipped %d/%zu steps (%.2fx estimated speedup)",
+                     spectrum_state.total_steps_skipped, total_steps, speedup);
+        }
+
         if (inverse_noise_scaling) {
             x = denoiser->inverse_noise_scaling(sigmas[sigmas.size() - 1], x);
         }
@@ -2515,14 +2597,14 @@ public:
         tile_size_y = get_tile_size(params.tile_size_y, params.rel_size_y, latent_y);
     }
 
-    ggml_tensor* vae_encode(ggml_context* work_ctx, ggml_tensor* x, bool encode_video = false) {
+    ggml_tensor* vae_encode(ggml_context* work_ctx, ggml_tensor* x) {
         int64_t t0                 = ggml_time_ms();
         ggml_tensor* result        = nullptr;
         const int vae_scale_factor = get_vae_scale_factor();
         int64_t W                  = x->ne[0] / vae_scale_factor;
         int64_t H                  = x->ne[1] / vae_scale_factor;
         int64_t C                  = get_latent_channel();
-        if (vae_tiling_params.enabled && !encode_video) {
+        if (vae_tiling_params.enabled) {
             // TODO wan2.2 vae support?
             int64_t ne2;
             int64_t ne3;
@@ -2550,7 +2632,7 @@ public:
 
         if (!use_tiny_autoencoder) {
             process_vae_input_tensor(x);
-            if (vae_tiling_params.enabled && !encode_video) {
+            if (vae_tiling_params.enabled) {
                 float tile_overlap;
                 int tile_size_x, tile_size_y;
                 // multiply tile size for encode to keep the compute buffer size consistent
@@ -2561,18 +2643,18 @@ public:
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
                     return first_stage_model->compute(n_threads, in, false, &out, work_ctx);
                 };
-                sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, on_tiling);
+                sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, circular_x, circular_y, on_tiling);
             } else {
                 first_stage_model->compute(n_threads, x, false, &result, work_ctx);
             }
             first_stage_model->free_compute_buffer();
         } else {
-            if (vae_tiling_params.enabled && !encode_video) {
+            if (vae_tiling_params.enabled) {
                 // split latent in 32x32 tiles and compute in several steps
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
                     return tae_first_stage->compute(n_threads, in, false, &out, nullptr);
                 };
-                sd_tiling(x, result, vae_scale_factor, 64, 0.5f, on_tiling);
+                sd_tiling(x, result, vae_scale_factor, 64, 0.5f, circular_x, circular_y, on_tiling);
             } else {
                 tae_first_stage->compute(n_threads, x, false, &result, work_ctx);
             }
@@ -2634,7 +2716,7 @@ public:
         } else {
             latent = gaussian_latent_sample(work_ctx, vae_output);
         }
-        if (!use_tiny_autoencoder) {
+        if (!use_tiny_autoencoder && version != VERSION_SD1_PIX2PIX) {
             process_latent_in(latent);
         }
         if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
@@ -2643,8 +2725,8 @@ public:
         return latent;
     }
 
-    ggml_tensor* encode_first_stage(ggml_context* work_ctx, ggml_tensor* x, bool encode_video = false) {
-        ggml_tensor* vae_output = vae_encode(work_ctx, x, encode_video);
+    ggml_tensor* encode_first_stage(ggml_context* work_ctx, ggml_tensor* x) {
+        ggml_tensor* vae_output = vae_encode(work_ctx, x);
         return get_first_stage_encoding(work_ctx, vae_output);
     }
 
@@ -2691,7 +2773,7 @@ public:
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
                     return first_stage_model->compute(n_threads, in, true, &out, nullptr);
                 };
-                sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, on_tiling);
+                sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, circular_x, circular_y, on_tiling);
             } else {
                 if (!first_stage_model->compute(n_threads, x, true, &result, work_ctx)) {
                     LOG_ERROR("Failed to decode latetnts");
@@ -2707,7 +2789,7 @@ public:
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
                     return tae_first_stage->compute(n_threads, in, true, &out);
                 };
-                sd_tiling(x, result, vae_scale_factor, 64, 0.5f, on_tiling);
+                sd_tiling(x, result, vae_scale_factor, 64, 0.5f, circular_x, circular_y, on_tiling);
             } else {
                 if (!tae_first_stage->compute(n_threads, x, true, &result)) {
                     LOG_ERROR("Failed to decode latetnts");
@@ -2914,7 +2996,7 @@ enum lora_apply_mode_t str_to_lora_apply_mode(const char* str) {
 void sd_cache_params_init(sd_cache_params_t* cache_params) {
     *cache_params                             = {};
     cache_params->mode                        = SD_CACHE_DISABLED;
-    cache_params->reuse_threshold             = 1.0f;
+    cache_params->reuse_threshold             = INFINITY;
     cache_params->start_percent               = 0.15f;
     cache_params->end_percent                 = 0.95f;
     cache_params->error_decay_rate            = 1.0f;
@@ -2930,6 +3012,13 @@ void sd_cache_params_init(sd_cache_params_t* cache_params) {
     cache_params->taylorseer_skip_interval    = 1;
     cache_params->scm_mask                    = nullptr;
     cache_params->scm_policy_dynamic          = true;
+    cache_params->spectrum_w                  = 0.40f;
+    cache_params->spectrum_m                  = 3;
+    cache_params->spectrum_lam                = 1.0f;
+    cache_params->spectrum_window_size        = 2;
+    cache_params->spectrum_flex_window        = 0.50f;
+    cache_params->spectrum_warmup_steps       = 4;
+    cache_params->spectrum_stop_percent       = 0.9f;
 }
 
 void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
@@ -3153,7 +3242,7 @@ char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params) {
     snprintf(buf + strlen(buf), 4096 - strlen(buf),
              "cache: %s (threshold=%.3f, start=%.2f, end=%.2f)\n",
              cache_mode_str,
-             sd_img_gen_params->cache.reuse_threshold,
+             get_cache_reuse_threshold(sd_img_gen_params->cache),
              sd_img_gen_params->cache.start_percent,
              sd_img_gen_params->cache.end_percent);
     free(sample_params_str);
@@ -3510,8 +3599,9 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
 
 sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_gen_params) {
     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 width  = sd_img_gen_params->width;
+    int height = sd_img_gen_params->height;
 
     int vae_scale_factor            = sd_ctx->sd->get_vae_scale_factor();
     int diffusion_model_down_factor = sd_ctx->sd->get_diffusion_model_down_factor();
@@ -3525,6 +3615,40 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
         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);
     }
 
+    bool circular_x = sd_ctx->sd->circular_x;
+    bool circular_y = sd_ctx->sd->circular_y;
+
+    if (!sd_img_gen_params->vae_tiling_params.enabled) {
+        if (sd_ctx->sd->first_stage_model) {
+            sd_ctx->sd->first_stage_model->set_circular_axes(sd_ctx->sd->circular_x, sd_ctx->sd->circular_y);
+        }
+        if (sd_ctx->sd->tae_first_stage) {
+            sd_ctx->sd->tae_first_stage->set_circular_axes(sd_ctx->sd->circular_x, sd_ctx->sd->circular_y);
+        }
+    } else {
+        int tile_size_x, tile_size_y;
+        float _overlap;
+        int latent_size_x = width / sd_ctx->sd->get_vae_scale_factor();
+        int latent_size_y = height / sd_ctx->sd->get_vae_scale_factor();
+        sd_ctx->sd->get_tile_sizes(tile_size_x, tile_size_y, _overlap, sd_img_gen_params->vae_tiling_params, latent_size_x, latent_size_y);
+
+        // force disable circular padding for vae if tiling is enabled unless latent is smaller than tile size
+        // otherwise it will cause artifacts at the edges of the tiles
+        sd_ctx->sd->circular_x = sd_ctx->sd->circular_x && (tile_size_x >= latent_size_x);
+        sd_ctx->sd->circular_y = sd_ctx->sd->circular_y && (tile_size_y >= latent_size_y);
+
+        if (sd_ctx->sd->first_stage_model) {
+            sd_ctx->sd->first_stage_model->set_circular_axes(sd_ctx->sd->circular_x, sd_ctx->sd->circular_y);
+        }
+        if (sd_ctx->sd->tae_first_stage) {
+            sd_ctx->sd->tae_first_stage->set_circular_axes(sd_ctx->sd->circular_x, sd_ctx->sd->circular_y);
+        }
+
+        // disable circular tiling if it's enabled for the VAE
+        sd_ctx->sd->circular_x = circular_x && (tile_size_x < latent_size_x);
+        sd_ctx->sd->circular_y = circular_y && (tile_size_y < latent_size_y);
+    }
+
     LOG_DEBUG("generate_image %dx%d", width, height);
     if (sd_ctx == nullptr || sd_img_gen_params == nullptr) {
         return nullptr;
@@ -3794,6 +3918,10 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
                                                         denoise_mask,
                                                         &sd_img_gen_params->cache);
 
+    // restore circular params
+    sd_ctx->sd->circular_x = circular_x;
+    sd_ctx->sd->circular_y = circular_y;
+
     size_t t2 = ggml_time_ms();
 
     LOG_INFO("generate_image completed in %.2fs", (t2 - t0) * 1.0f / 1000);

src/ucache.hpp

@@ -19,6 +19,7 @@ struct UCacheConfig {
     bool adaptive_threshold     = true;
     float early_step_multiplier = 0.5f;
     float late_step_multiplier  = 1.5f;
+    float relative_norm_gain    = 1.6f;
     bool reset_error_on_compute = true;
 };
 
@@ -45,14 +46,16 @@ struct UCacheState {
     bool has_output_prev_norm             = false;
     bool has_relative_transformation_rate = false;
     float relative_transformation_rate    = 0.0f;
-    float cumulative_change_rate          = 0.0f;
     float last_input_change               = 0.0f;
     bool has_last_input_change            = false;
+    float output_change_ema               = 0.0f;
+    bool has_output_change_ema            = false;
     int total_steps_skipped               = 0;
     int current_step_index                = -1;
     int steps_computed_since_active       = 0;
+    int expected_total_steps              = 0;
+    int consecutive_skipped_steps         = 0;
     float accumulated_error               = 0.0f;
-    float reference_output_norm           = 0.0f;
 
     struct BlockMetrics {
         float sum_transformation_rate = 0.0f;
@@ -106,14 +109,16 @@ struct UCacheState {
         has_output_prev_norm             = false;
         has_relative_transformation_rate = false;
         relative_transformation_rate     = 0.0f;
-        cumulative_change_rate           = 0.0f;
         last_input_change                = 0.0f;
         has_last_input_change            = false;
+        output_change_ema                = 0.0f;
+        has_output_change_ema            = false;
         total_steps_skipped              = 0;
         current_step_index               = -1;
         steps_computed_since_active      = 0;
+        expected_total_steps             = 0;
+        consecutive_skipped_steps        = 0;
         accumulated_error                = 0.0f;
-        reference_output_norm            = 0.0f;
         block_metrics.reset();
         total_active_steps = 0;
     }
@@ -133,7 +138,8 @@ struct UCacheState {
         if (!initialized || sigmas.size() < 2) {
             return;
         }
-        size_t n_steps = sigmas.size() - 1;
+        size_t n_steps       = sigmas.size() - 1;
+        expected_total_steps = static_cast<int>(n_steps);
 
         size_t start_step = static_cast<size_t>(config.start_percent * n_steps);
         size_t end_step   = static_cast<size_t>(config.end_percent * n_steps);
@@ -207,11 +213,15 @@ struct UCacheState {
         }
 
         int effective_total = estimated_total_steps;
+        if (effective_total <= 0) {
+            effective_total = expected_total_steps;
+        }
         if (effective_total <= 0) {
             effective_total = std::max(20, steps_computed_since_active * 2);
         }
 
         float progress = (effective_total > 0) ? (static_cast<float>(steps_computed_since_active) / effective_total) : 0.0f;
+        progress       = std::max(0.0f, std::min(1.0f, progress));
 
         float multiplier = 1.0f;
         if (progress < 0.2f) {
@@ -309,17 +319,31 @@ struct UCacheState {
 
         if (has_output_prev_norm && has_relative_transformation_rate &&
             last_input_change > 0.0f && output_prev_norm > 0.0f) {
-            float approx_output_change_rate = (relative_transformation_rate * last_input_change) / output_prev_norm;
-            accumulated_error               = accumulated_error * config.error_decay_rate + approx_output_change_rate;
+            float approx_output_change = relative_transformation_rate * last_input_change;
+            float approx_output_change_rate;
+            if (config.use_relative_threshold) {
+                float base_scale          = std::max(output_prev_norm, 1e-6f);
+                float dyn_scale           = has_output_change_ema
+                                                ? std::max(output_change_ema * std::max(1.0f, config.relative_norm_gain), 1e-6f)
+                                                : base_scale;
+                float scale               = std::sqrt(base_scale * dyn_scale);
+                approx_output_change_rate = approx_output_change / scale;
+            } else {
+                approx_output_change_rate = approx_output_change;
+            }
+            // Increase estimated error with skip horizon to avoid long extrapolation streaks
+            approx_output_change_rate *= (1.0f + 0.50f * consecutive_skipped_steps);
+            accumulated_error = accumulated_error * config.error_decay_rate + approx_output_change_rate;
 
             float effective_threshold = get_adaptive_threshold();
-            if (config.use_relative_threshold && reference_output_norm > 0.0f) {
-                effective_threshold = effective_threshold * reference_output_norm;
+            if (!config.use_relative_threshold && output_prev_norm > 0.0f) {
+                effective_threshold = effective_threshold * output_prev_norm;
             }
 
             if (accumulated_error < effective_threshold) {
                 skip_current_step = true;
                 total_steps_skipped++;
+                consecutive_skipped_steps++;
                 apply_cache(cond, input, output);
                 return true;
             } else if (config.reset_error_on_compute) {
@@ -340,6 +364,8 @@ struct UCacheState {
         if (cond != anchor_condition) {
             return;
         }
+        steps_computed_since_active++;
+        consecutive_skipped_steps = 0;
 
         size_t ne      = static_cast<size_t>(ggml_nelements(input));
         float* in_data = (float*)input->data;
@@ -359,6 +385,14 @@ struct UCacheState {
                 output_change /= static_cast<float>(ne);
             }
         }
+        if (std::isfinite(output_change) && output_change > 0.0f) {
+            if (!has_output_change_ema) {
+                output_change_ema     = output_change;
+                has_output_change_ema = true;
+            } else {
+                output_change_ema = 0.8f * output_change_ema + 0.2f * output_change;
+            }
+        }
 
         prev_output.resize(ne);
         for (size_t i = 0; i < ne; ++i) {
@@ -373,10 +407,6 @@ struct UCacheState {
         output_prev_norm     = (ne > 0) ? (mean_abs / static_cast<float>(ne)) : 0.0f;
         has_output_prev_norm = output_prev_norm > 0.0f;
 
-        if (reference_output_norm == 0.0f) {
-            reference_output_norm = output_prev_norm;
-        }
-
         if (has_last_input_change && last_input_change > 0.0f && output_change > 0.0f) {
             float rate = output_change / last_input_change;
             if (std::isfinite(rate)) {

src/upscaler.cpp

@@ -92,7 +92,8 @@ struct UpscalerGGML {
             return esrgan_upscaler->compute(n_threads, in, &out);
         };
         int64_t t0 = ggml_time_ms();
-        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, on_tiling);
+        // TODO: circular upscaling?
+        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, false, false, on_tiling);
         esrgan_upscaler->free_compute_buffer();
         ggml_ext_tensor_clamp_inplace(upscaled, 0.f, 1.f);
         uint8_t* upscaled_data = ggml_tensor_to_sd_image(upscaled);