fix: correct double increment on flow denoisers sigma calculations (#1372)

* inpaint: get max pixel max instead of single sample

* inpaint: masked diffusion for inpainting models with inflated mask

* refactor tensor interpolate nearest-like reduction paths and generalize max_pool_2d

---------

Co-authored-by: leejet <leejet714@gmail.com>

.github/workflows/build.yml

@@ -64,7 +64,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Dependencies
         id: depends
@@ -127,7 +127,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Dependencies
         id: depends
@@ -205,7 +205,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Get commit hash
         id: commit
@@ -239,6 +239,7 @@ jobs:
         id: build-push
         uses: docker/build-push-action@v6
         with:
+          context: .
           platforms: linux/amd64
           push: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
           file: Dockerfile.${{ matrix.variant }}
@@ -264,7 +265,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Dependencies
         id: depends
@@ -345,7 +346,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Install cuda-toolkit
         id: cuda-toolkit
@@ -460,7 +461,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Cache ROCm Installation
         id: cache-rocm
@@ -573,7 +574,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Free disk space
         run: |

.gitmodules

@@ -4,3 +4,9 @@
 [submodule "examples/server/frontend"]
 	path = examples/server/frontend
 	url = https://github.com/leejet/stable-ui.git
+[submodule "thirdparty/libwebp"]
+	path = thirdparty/libwebp
+	url = https://github.com/webmproject/libwebp.git
+[submodule "thirdparty/libwebm"]
+	path = thirdparty/libwebm
+	url = https://github.com/webmproject/libwebm.git

CMakeLists.txt

@@ -22,6 +22,26 @@ else()
     set(SD_STANDALONE OFF)
 endif()
 
+set(SD_SUBMODULE_WEBP FALSE)
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/libwebp/CMakeLists.txt")
+    set(SD_SUBMODULE_WEBP TRUE)
+endif()
+if(SD_SUBMODULE_WEBP)
+    set(SD_WEBP_DEFAULT ON)
+else()
+    set(SD_WEBP_DEFAULT ${SD_USE_SYSTEM_WEBP})
+endif()
+
+set(SD_SUBMODULE_WEBM FALSE)
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/libwebm/CMakeLists.txt")
+    set(SD_SUBMODULE_WEBM TRUE)
+endif()
+if(SD_SUBMODULE_WEBM)
+    set(SD_WEBM_DEFAULT ON)
+else()
+    set(SD_WEBM_DEFAULT ${SD_USE_SYSTEM_WEBM})
+endif()
+
 #
 # Option list
 #
@@ -29,6 +49,10 @@ endif()
 # general
 #option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
 option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
+option(SD_WEBP                       "sd: enable WebP image I/O support" ${SD_WEBP_DEFAULT})
+option(SD_USE_SYSTEM_WEBP            "sd: link against system libwebp" OFF)
+option(SD_WEBM                       "sd: enable WebM video output support" ${SD_WEBM_DEFAULT})
+option(SD_USE_SYSTEM_WEBM            "sd: link against system libwebm" OFF)
 option(SD_CUDA                       "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
@@ -77,6 +101,55 @@ if(SD_MUSA)
     add_definitions(-DSD_USE_CUDA)
 endif()
 
+if(SD_WEBP)
+    if(NOT SD_SUBMODULE_WEBP AND NOT SD_USE_SYSTEM_WEBP)
+        message(FATAL_ERROR "WebP support enabled but no source found.
+          Either initialize the submodule:\n  git submodule update --init thirdparty/libwebp\n\n"
+          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBP=ON")
+    endif()
+    if(SD_USE_SYSTEM_WEBP)
+        find_package(WebP REQUIRED)
+        add_library(webp ALIAS WebP::webp)
+        # libwebp CMake target naming is not consistent across versions/distros.
+        # Some export WebP::libwebpmux, others export WebP::webpmux.
+        if(TARGET WebP::libwebpmux)
+            add_library(libwebpmux ALIAS WebP::libwebpmux)
+        elseif(TARGET WebP::webpmux)
+            add_library(libwebpmux ALIAS WebP::webpmux)
+        else()
+            message(FATAL_ERROR
+                "Could not find a compatible webpmux target in system WebP package. "
+                "Expected WebP::libwebpmux or WebP::webpmux."
+            )
+        endif()
+    endif()
+endif()
+
+if(SD_WEBM)
+    if(NOT SD_WEBP)
+        message(FATAL_ERROR "SD_WEBM requires SD_WEBP because WebM output reuses libwebp VP8 encoding.")
+    endif()
+    if(NOT SD_SUBMODULE_WEBM AND NOT SD_USE_SYSTEM_WEBM)
+        message(FATAL_ERROR "WebM support enabled but no source found.
+          Either initialize the submodule:\n  git submodule update --init thirdparty/libwebm\n\n"
+          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBM=ON")
+    endif()
+    if(SD_USE_SYSTEM_WEBM)
+        find_path(WEBM_INCLUDE_DIR
+            NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
+            PATH_SUFFIXES webm
+            REQUIRED)
+        find_library(WEBM_LIBRARY
+            NAMES webm libwebm
+            REQUIRED)
+
+        add_library(webm UNKNOWN IMPORTED)
+        set_target_properties(webm PROPERTIES
+            IMPORTED_LOCATION "${WEBM_LIBRARY}"
+            INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+    endif()
+endif()
+
 set(SD_LIB stable-diffusion)
 
 file(GLOB SD_LIB_SOURCES

docs/build.md

@@ -16,6 +16,26 @@ git submodule init
 git submodule update
 ```
 
+## WebP and WebM Support in Examples
+
+The example applications (`examples/cli` and `examples/server`) use `libwebp` to support WebP image I/O, and `examples/cli` can also use `libwebm` for `.webm` video output. Both are enabled by default. WebM output currently reuses `libwebp` to encode each frame as VP8 before muxing with `libwebm`.
+
+If you do not want WebP/WebM support, you can disable them at configure time:
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_WEBP=OFF -DSD_WEBM=OFF
+cmake --build . --config Release
+```
+
+If the submodules are not available, you can also link against system packages instead:
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_USE_SYSTEM_WEBP=ON -DSD_USE_SYSTEM_WEBM=ON
+cmake --build . --config Release
+```
+
 ## Build (CPU only)
 
 If you don't have a GPU or CUDA installed, you can build a CPU-only version.

examples/cli/CMakeLists.txt

@@ -1,6 +1,20 @@
 set(TARGET sd-cli)
 
-add_executable(${TARGET} main.cpp)
+add_executable(${TARGET}
+    ../common/common.cpp
+    ../common/log.cpp
+    ../common/media_io.cpp
+    image_metadata.cpp
+    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)
+target_link_libraries(${TARGET} PRIVATE stable-diffusion zip ${CMAKE_THREAD_LIBS_INIT})
+if(SD_WEBP)
+    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBP)
+    target_link_libraries(${TARGET} PRIVATE webp libwebpmux)
+endif()
+if(SD_WEBM)
+    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBM)
+    target_link_libraries(${TARGET} PRIVATE webm)
+endif()
+target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)

examples/cli/README.md

@@ -5,18 +5,23 @@ 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)
-  --preview-path <string>     path to write preview image to (default: ./preview.png)
+                              ./output.png) (eg. output_%03d.png). For video generation, single-file outputs support .avi, .webm, and animated .webp
+  --preview-path <string>     path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp
   --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)
+  --image <string>            path to the image to inspect (for metadata mode)
+  --metadata-format <string>  metadata output format, one of [text, json] (default: text)
   --canny                     apply canny preprocessor (edge detection)
   --convert-name              convert tensor name (for convert mode)
   -v, --verbose               print extra info
   --color                     colors the logging tags according to level
   --taesd-preview-only        prevents usage of taesd for decoding the final image. (for use with --preview tae)
   --preview-noisy             enables previewing noisy inputs of the models rather than the denoised outputs
-  -M, --mode                  run mode, one of [img_gen, vid_gen, upscale, convert], default: img_gen
+  --metadata-raw              include raw hex previews for unparsed metadata payloads
+  --metadata-brief            truncate long metadata text values in text output
+  --metadata-all              include structural/container entries such as IHDR, IDAT, and non-metadata JPEG segments
+  -M, --mode                  run mode, one of [img_gen, vid_gen, upscale, convert, metadata], default: img_gen
   --preview                   preview method. must be one of the following [none, proj, tae, vae] (default is none)
   -h, --help                  show this help message and exit
 
@@ -109,7 +114,7 @@ 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 and TCD (default: 0)
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)
   --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)
@@ -117,7 +122,7 @@ 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 and TCD (default: 0)
+  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)
   --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
@@ -148,3 +153,12 @@ Generation Options:
   --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'
 ```
+
+Metadata mode inspects PNG/JPEG container metadata without loading any model:
+
+```bash
+./bin/sd-cli -M metadata --image ./output.png
+./bin/sd-cli -M metadata --image ./output.jpg --metadata-format json
+./bin/sd-cli -M metadata --image ./output.png --metadata-raw
+./bin/sd-cli -M metadata --image ./output.png --metadata-all
+```

examples/cli/avi_writer.h

@@ -1,217 +0,0 @@
-#ifndef __AVI_WRITER_H__
-#define __AVI_WRITER_H__
-
-#include <cstdint>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-
-#include "stable-diffusion.h"
-
-#ifndef INCLUDE_STB_IMAGE_WRITE_H
-#include "stb_image_write.h"
-#endif
-
-typedef struct {
-    uint32_t offset;
-    uint32_t size;
-} avi_index_entry;
-
-// Write 32-bit little-endian integer
-void write_u32_le(FILE* f, uint32_t val) {
-    fwrite(&val, 4, 1, f);
-}
-
-// Write 16-bit little-endian integer
-void write_u16_le(FILE* f, uint16_t val) {
-    fwrite(&val, 2, 1, f);
-}
-
-/**
- * Create an MJPG AVI file from an array of sd_image_t images.
- * Images are encoded to JPEG using stb_image_write.
- *
- * @param filename Output AVI file name.
- * @param images Array of input images.
- * @param num_images Number of images in the array.
- * @param fps Frames per second for the video.
- * @param quality JPEG quality (0-100).
- * @return 0 on success, -1 on failure.
- */
-int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality = 90) {
-    if (num_images == 0) {
-        fprintf(stderr, "Error: Image array is empty.\n");
-        return -1;
-    }
-
-    FILE* f = fopen(filename, "wb");
-    if (!f) {
-        perror("Error opening file for writing");
-        return -1;
-    }
-
-    uint32_t width    = images[0].width;
-    uint32_t height   = images[0].height;
-    uint32_t channels = images[0].channel;
-    if (channels != 3 && channels != 4) {
-        fprintf(stderr, "Error: Unsupported channel count: %u\n", channels);
-        fclose(f);
-        return -1;
-    }
-
-    // --- RIFF AVI Header ---
-    fwrite("RIFF", 4, 1, f);
-    long riff_size_pos = ftell(f);
-    write_u32_le(f, 0);  // Placeholder for file size
-    fwrite("AVI ", 4, 1, f);
-
-    // 'hdrl' LIST (header list)
-    fwrite("LIST", 4, 1, f);
-    write_u32_le(f, 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40);
-    fwrite("hdrl", 4, 1, f);
-
-    // 'avih' chunk (AVI main header)
-    fwrite("avih", 4, 1, f);
-    write_u32_le(f, 56);
-    write_u32_le(f, 1000000 / fps);       // Microseconds per frame
-    write_u32_le(f, 0);                   // Max bytes per second
-    write_u32_le(f, 0);                   // Padding granularity
-    write_u32_le(f, 0x110);               // Flags (HASINDEX | ISINTERLEAVED)
-    write_u32_le(f, num_images);          // Total frames
-    write_u32_le(f, 0);                   // Initial frames
-    write_u32_le(f, 1);                   // Number of streams
-    write_u32_le(f, width * height * 3);  // Suggested buffer size
-    write_u32_le(f, width);
-    write_u32_le(f, height);
-    write_u32_le(f, 0);  // Reserved
-    write_u32_le(f, 0);  // Reserved
-    write_u32_le(f, 0);  // Reserved
-    write_u32_le(f, 0);  // Reserved
-
-    // 'strl' LIST (stream list)
-    fwrite("LIST", 4, 1, f);
-    write_u32_le(f, 4 + 8 + 56 + 8 + 40);
-    fwrite("strl", 4, 1, f);
-
-    // 'strh' chunk (stream header)
-    fwrite("strh", 4, 1, f);
-    write_u32_le(f, 56);
-    fwrite("vids", 4, 1, f);              // Stream type: video
-    fwrite("MJPG", 4, 1, f);              // Codec: Motion JPEG
-    write_u32_le(f, 0);                   // Flags
-    write_u16_le(f, 0);                   // Priority
-    write_u16_le(f, 0);                   // Language
-    write_u32_le(f, 0);                   // Initial frames
-    write_u32_le(f, 1);                   // Scale
-    write_u32_le(f, fps);                 // Rate
-    write_u32_le(f, 0);                   // Start
-    write_u32_le(f, num_images);          // Length
-    write_u32_le(f, width * height * 3);  // Suggested buffer size
-    write_u32_le(f, (uint32_t)-1);        // Quality
-    write_u32_le(f, 0);                   // Sample size
-    write_u16_le(f, 0);                   // rcFrame.left
-    write_u16_le(f, 0);                   // rcFrame.top
-    write_u16_le(f, 0);                   // rcFrame.right
-    write_u16_le(f, 0);                   // rcFrame.bottom
-
-    // 'strf' chunk (stream format: BITMAPINFOHEADER)
-    fwrite("strf", 4, 1, f);
-    write_u32_le(f, 40);
-    write_u32_le(f, 40);  // biSize
-    write_u32_le(f, width);
-    write_u32_le(f, height);
-    write_u16_le(f, 1);                   // biPlanes
-    write_u16_le(f, 24);                  // biBitCount
-    fwrite("MJPG", 4, 1, f);              // biCompression (FOURCC)
-    write_u32_le(f, width * height * 3);  // biSizeImage
-    write_u32_le(f, 0);                   // XPelsPerMeter
-    write_u32_le(f, 0);                   // YPelsPerMeter
-    write_u32_le(f, 0);                   // Colors used
-    write_u32_le(f, 0);                   // Colors important
-
-    // 'movi' LIST (video frames)
-    // long movi_list_pos = ftell(f);
-    fwrite("LIST", 4, 1, f);
-    long movi_size_pos = ftell(f);
-    write_u32_le(f, 0);  // Placeholder for movi size
-    fwrite("movi", 4, 1, f);
-
-    avi_index_entry* index = (avi_index_entry*)malloc(sizeof(avi_index_entry) * num_images);
-    if (!index) {
-        fclose(f);
-        return -1;
-    }
-
-    // Encode and write each frame as JPEG
-    struct {
-        uint8_t* buf;
-        size_t size;
-    } jpeg_data;
-
-    for (int i = 0; i < num_images; i++) {
-        jpeg_data.buf  = nullptr;
-        jpeg_data.size = 0;
-
-        // Callback function to collect JPEG data into memory
-        auto write_to_buf = [](void* context, void* data, int size) {
-            auto jd = (decltype(jpeg_data)*)context;
-            jd->buf = (uint8_t*)realloc(jd->buf, jd->size + size);
-            memcpy(jd->buf + jd->size, data, size);
-            jd->size += size;
-        };
-
-        // Encode to JPEG in memory
-        stbi_write_jpg_to_func(
-            write_to_buf,
-            &jpeg_data,
-            images[i].width,
-            images[i].height,
-            channels,
-            images[i].data,
-            quality);
-
-        // Write '00dc' chunk (video frame)
-        fwrite("00dc", 4, 1, f);
-        write_u32_le(f, (uint32_t)jpeg_data.size);
-        index[i].offset = ftell(f) - 8;
-        index[i].size   = (uint32_t)jpeg_data.size;
-        fwrite(jpeg_data.buf, 1, jpeg_data.size, f);
-
-        // Align to even byte size
-        if (jpeg_data.size % 2)
-            fputc(0, f);
-
-        free(jpeg_data.buf);
-    }
-
-    // Finalize 'movi' size
-    long cur_pos   = ftell(f);
-    long movi_size = cur_pos - movi_size_pos - 4;
-    fseek(f, movi_size_pos, SEEK_SET);
-    write_u32_le(f, movi_size);
-    fseek(f, cur_pos, SEEK_SET);
-
-    // Write 'idx1' index
-    fwrite("idx1", 4, 1, f);
-    write_u32_le(f, num_images * 16);
-    for (int i = 0; i < num_images; i++) {
-        fwrite("00dc", 4, 1, f);
-        write_u32_le(f, 0x10);
-        write_u32_le(f, index[i].offset);
-        write_u32_le(f, index[i].size);
-    }
-
-    // Finalize RIFF size
-    cur_pos        = ftell(f);
-    long file_size = cur_pos - riff_size_pos - 4;
-    fseek(f, riff_size_pos, SEEK_SET);
-    write_u32_le(f, file_size);
-    fseek(f, cur_pos, SEEK_SET);
-
-    fclose(f);
-    free(index);
-
-    return 0;
-}
-
-#endif  // __AVI_WRITER_H__

examples/cli/image_metadata.h

@@ -0,0 +1,21 @@
+#pragma once
+
+#include <iosfwd>
+#include <string>
+
+enum class MetadataOutputFormat {
+    TEXT,
+    JSON,
+};
+
+struct MetadataReadOptions {
+    MetadataOutputFormat output_format = MetadataOutputFormat::TEXT;
+    bool include_raw                   = false;
+    bool brief                         = false;
+    bool include_structural            = false;
+};
+
+bool print_image_metadata(const std::string& image_path,
+                          const MetadataReadOptions& options,
+                          std::ostream& out,
+                          std::string& error);

examples/cli/main.cpp

@@ -15,9 +15,12 @@
 // #include "preprocessing.hpp"
 #include "stable-diffusion.h"
 
-#include "common/common.hpp"
+#include "common/common.h"
+#include "common/media_io.h"
+#include "common/resource_owners.hpp"
+#include "image_metadata.h"
 
-#include "avi_writer.h"
+namespace fs = std::filesystem;
 
 const char* previews_str[] = {
     "none",
@@ -32,6 +35,8 @@ struct SDCliParams {
     SDMode mode             = IMG_GEN;
     std::string output_path = "output.png";
     int output_begin_idx    = -1;
+    std::string image_path;
+    std::string metadata_format = "text";
 
     bool verbose          = false;
     bool canny_preprocess = false;
@@ -44,6 +49,9 @@ struct SDCliParams {
     bool taesd_preview       = false;
     bool preview_noisy       = false;
     bool color               = false;
+    bool metadata_raw        = false;
+    bool metadata_brief      = false;
+    bool metadata_all        = false;
 
     bool normal_exit = false;
 
@@ -53,11 +61,19 @@ struct SDCliParams {
         options.string_options = {
             {"-o",
              "--output",
-             "path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png)",
+             "path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs support .avi, .webm, and animated .webp",
              &output_path},
+            {"",
+             "--image",
+             "path to the image to inspect (for metadata mode)",
+             &image_path},
+            {"",
+             "--metadata-format",
+             "metadata output format, one of [text, json] (default: text)",
+             &metadata_format},
             {"",
              "--preview-path",
-             "path to write preview image to (default: ./preview.png)",
+             "path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp",
              &preview_path},
         };
 
@@ -97,6 +113,18 @@ struct SDCliParams {
              "--preview-noisy",
              "enables previewing noisy inputs of the models rather than the denoised outputs",
              true, &preview_noisy},
+            {"",
+             "--metadata-raw",
+             "include raw hex previews for unparsed metadata payloads",
+             true, &metadata_raw},
+            {"",
+             "--metadata-brief",
+             "truncate long metadata text values in text output",
+             true, &metadata_brief},
+            {"",
+             "--metadata-all",
+             "include structural/container entries such as IHDR, IDAT, and non-metadata JPEG segments",
+             true, &metadata_all},
 
         };
 
@@ -149,7 +177,7 @@ struct SDCliParams {
         options.manual_options = {
             {"-M",
              "--mode",
-             "run mode, one of [img_gen, vid_gen, upscale, convert], default: img_gen",
+             "run mode, one of [img_gen, vid_gen, upscale, convert, metadata], default: img_gen",
              on_mode_arg},
             {"",
              "--preview",
@@ -165,7 +193,7 @@ struct SDCliParams {
     };
 
     bool process_and_check() {
-        if (output_path.length() == 0) {
+        if (mode != METADATA && output_path.length() == 0) {
             LOG_ERROR("error: the following arguments are required: output_path");
             return false;
         }
@@ -174,6 +202,16 @@ struct SDCliParams {
             if (output_path == "output.png") {
                 output_path = "output.gguf";
             }
+        } else if (mode == METADATA) {
+            if (image_path.empty()) {
+                LOG_ERROR("error: metadata mode needs an image path (--image)");
+                return false;
+            }
+            if (metadata_format != "text" && metadata_format != "json") {
+                LOG_ERROR("error: invalid metadata format %s, must be one of [text, json]",
+                          metadata_format.c_str());
+                return false;
+            }
         }
         return true;
     }
@@ -183,6 +221,8 @@ struct SDCliParams {
         oss << "SDCliParams {\n"
             << "  mode: " << modes_str[mode] << ",\n"
             << "  output_path: \"" << output_path << "\",\n"
+            << "  image_path: \"" << image_path << "\",\n"
+            << "  metadata_format: \"" << metadata_format << "\",\n"
             << "  verbose: " << (verbose ? "true" : "false") << ",\n"
             << "  color: " << (color ? "true" : "false") << ",\n"
             << "  canny_preprocess: " << (canny_preprocess ? "true" : "false") << ",\n"
@@ -192,7 +232,10 @@ struct SDCliParams {
             << "  preview_path: \"" << preview_path << "\",\n"
             << "  preview_fps: " << preview_fps << ",\n"
             << "  taesd_preview: " << (taesd_preview ? "true" : "false") << ",\n"
-            << "  preview_noisy: " << (preview_noisy ? "true" : "false") << "\n"
+            << "  preview_noisy: " << (preview_noisy ? "true" : "false") << ",\n"
+            << "  metadata_raw: " << (metadata_raw ? "true" : "false") << ",\n"
+            << "  metadata_brief: " << (metadata_brief ? "true" : "false") << ",\n"
+            << "  metadata_all: " << (metadata_all ? "true" : "false") << "\n"
             << "}";
         return oss.str();
     }
@@ -217,9 +260,13 @@ void parse_args(int argc, const char** argv, SDCliParams& cli_params, SDContextP
         exit(cli_params.normal_exit ? 0 : 1);
     }
 
-    if (!cli_params.process_and_check() ||
-        !ctx_params.process_and_check(cli_params.mode) ||
-        !gen_params.process_and_check(cli_params.mode, ctx_params.lora_model_dir)) {
+    bool valid = cli_params.process_and_check();
+    if (valid && cli_params.mode != METADATA) {
+        valid = ctx_params.process_and_check(cli_params.mode) &&
+                gen_params.process_and_check(cli_params.mode, ctx_params.lora_model_dir);
+    }
+
+    if (!valid) {
         print_usage(argc, argv, options_vec);
         exit(1);
     }
@@ -231,7 +278,7 @@ void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
 }
 
 bool load_images_from_dir(const std::string dir,
-                          std::vector<sd_image_t>& images,
+                          SDImageVec& images,
                           int expected_width  = 0,
                           int expected_height = 0,
                           int max_image_num   = 0,
@@ -258,7 +305,7 @@ bool load_images_from_dir(const std::string dir,
         std::string ext  = entry.path().extension().string();
         std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
 
-        if (ext == ".jpg" || ext == ".jpeg" || ext == ".png" || ext == ".bmp") {
+        if (ext == ".jpg" || ext == ".jpeg" || ext == ".png" || ext == ".bmp" || ext == ".webp") {
             LOG_DEBUG("load image %zu from '%s'", images.size(), path.c_str());
             int width             = 0;
             int height            = 0;
@@ -273,7 +320,7 @@ bool load_images_from_dir(const std::string dir,
                               3,
                               image_buffer});
 
-            if (max_image_num > 0 && images.size() >= max_image_num) {
+            if (max_image_num > 0 && static_cast<int>(images.size()) >= max_image_num) {
                 break;
             }
         }
@@ -288,9 +335,17 @@ void step_callback(int step, int frame_count, sd_image_t* image, bool is_noisy,
     // is_noisy is set to true if the preview corresponds to noisy latents, false if it's denoised latents
     // unused in this app, it will either be always noisy or always denoised here
     if (frame_count == 1) {
-        stbi_write_png(cli_params->preview_path.c_str(), image->width, image->height, image->channel, image->data, 0);
+        if (!write_image_to_file(cli_params->preview_path,
+                                 image->data,
+                                 image->width,
+                                 image->height,
+                                 image->channel)) {
+            LOG_ERROR("save preview image to '%s' failed", cli_params->preview_path.c_str());
+        }
     } else {
-        create_mjpg_avi_from_sd_images(cli_params->preview_path.c_str(), image, frame_count, cli_params->preview_fps);
+        if (create_video_from_sd_images(cli_params->preview_path.c_str(), image, frame_count, cli_params->preview_fps) != 0) {
+            LOG_ERROR("save preview video to '%s' failed", cli_params->preview_path.c_str());
+        }
     }
 }
 
@@ -340,9 +395,13 @@ bool save_results(const SDCliParams& cli_params,
 
     std::string ext_lower = ext.string();
     std::transform(ext_lower.begin(), ext_lower.end(), ext_lower.begin(), ::tolower);
-    bool is_jpg = (ext_lower == ".jpg" || ext_lower == ".jpeg" || ext_lower == ".jpe");
+    const EncodedImageFormat output_format = encoded_image_format_from_path(out_path.string());
     if (!ext.empty()) {
-        if (is_jpg || ext_lower == ".png") {
+        if (output_format == EncodedImageFormat::JPEG ||
+            output_format == EncodedImageFormat::PNG ||
+            output_format == EncodedImageFormat::WEBP ||
+            ext_lower == ".avi" ||
+            ext_lower == ".webm") {
             base_path.replace_extension();
         }
     }
@@ -360,20 +419,15 @@ bool save_results(const SDCliParams& cli_params,
         std::string params = gen_params.embed_image_metadata
                                  ? get_image_params(ctx_params, gen_params, gen_params.seed + idx)
                                  : "";
-        int ok             = 0;
-        if (is_jpg) {
-            ok = stbi_write_jpg(path.string().c_str(), img.width, img.height, img.channel, img.data, 90, params.size() > 0 ? params.c_str() : nullptr);
-        } else {
-            ok = stbi_write_png(path.string().c_str(), img.width, img.height, img.channel, img.data, 0, params.size() > 0 ? params.c_str() : nullptr);
-        }
+        const bool ok      = write_image_to_file(path.string(), img.data, img.width, img.height, img.channel, params, 90);
         LOG_INFO("save result image %d to '%s' (%s)", idx, path.string().c_str(), ok ? "success" : "failure");
-        return ok != 0;
+        return ok;
     };
 
     int sucessful_reults = 0;
 
     if (std::regex_search(cli_params.output_path, format_specifier_regex)) {
-        if (!is_jpg && ext_lower != ".png")
+        if (output_format == EncodedImageFormat::UNKNOWN)
             ext = ".png";
         fs::path pattern = base_path;
         pattern += ext;
@@ -389,20 +443,20 @@ bool save_results(const SDCliParams& cli_params,
     }
 
     if (cli_params.mode == VID_GEN && num_results > 1) {
-        if (ext_lower != ".avi")
+        if (ext_lower != ".avi" && ext_lower != ".webp" && ext_lower != ".webm")
             ext = ".avi";
         fs::path video_path = base_path;
         video_path += ext;
-        if (create_mjpg_avi_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps) == 0) {
-            LOG_INFO("save result MJPG AVI video to '%s'", video_path.string().c_str());
+        if (create_video_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps) == 0) {
+            LOG_INFO("save result video to '%s'", video_path.string().c_str());
             return true;
         } else {
-            LOG_ERROR("Failed to save result MPG AVI video to '%s'", video_path.string().c_str());
+            LOG_ERROR("Failed to save result video to '%s'", video_path.string().c_str());
             return false;
         }
     }
 
-    if (!is_jpg && ext_lower != ".png")
+    if (output_format == EncodedImageFormat::UNKNOWN)
         ext = ".png";
 
     for (int i = 0; i < num_results; ++i) {
@@ -430,6 +484,27 @@ int main(int argc, const char* argv[]) {
     SDGenerationParams gen_params;
 
     parse_args(argc, argv, cli_params, ctx_params, gen_params);
+    sd_set_log_callback(sd_log_cb, (void*)&cli_params);
+    log_verbose = cli_params.verbose;
+    log_color   = cli_params.color;
+
+    if (cli_params.mode == METADATA) {
+        MetadataReadOptions options;
+        options.output_format      = cli_params.metadata_format == "json"
+                                         ? MetadataOutputFormat::JSON
+                                         : MetadataOutputFormat::TEXT;
+        options.include_raw        = cli_params.metadata_raw;
+        options.brief              = cli_params.metadata_brief;
+        options.include_structural = cli_params.metadata_all;
+
+        std::string error;
+        if (!print_image_metadata(cli_params.image_path, options, std::cout, error)) {
+            LOG_ERROR("%s", error.c_str());
+            return 1;
+        }
+        return 0;
+    }
+
     if (gen_params.video_frames > 4) {
         size_t last_dot_pos   = cli_params.preview_path.find_last_of(".");
         std::string base_path = cli_params.preview_path;
@@ -447,9 +522,6 @@ int main(int argc, const char* argv[]) {
     if (cli_params.preview_method == PREVIEW_PROJ)
         cli_params.preview_fps /= 4;
 
-    sd_set_log_callback(sd_log_cb, (void*)&cli_params);
-    log_verbose = cli_params.verbose;
-    log_color   = cli_params.color;
     sd_set_preview_callback(step_callback,
                             cli_params.preview_method,
                             cli_params.preview_interval,
@@ -485,39 +557,17 @@ int main(int argc, const char* argv[]) {
         }
     }
 
-    bool vae_decode_only     = true;
-    sd_image_t init_image    = {0, 0, 3, nullptr};
-    sd_image_t end_image     = {0, 0, 3, nullptr};
-    sd_image_t control_image = {0, 0, 3, nullptr};
-    sd_image_t mask_image    = {0, 0, 1, nullptr};
-    std::vector<sd_image_t> ref_images;
-    std::vector<sd_image_t> pmid_images;
-    std::vector<sd_image_t> control_frames;
-
-    auto release_all_resources = [&]() {
-        free(init_image.data);
-        free(end_image.data);
-        free(control_image.data);
-        free(mask_image.data);
-        for (auto image : ref_images) {
-            free(image.data);
-            image.data = nullptr;
-        }
-        ref_images.clear();
-        for (auto image : pmid_images) {
-            free(image.data);
-            image.data = nullptr;
-        }
-        pmid_images.clear();
-        for (auto image : control_frames) {
-            free(image.data);
-            image.data = nullptr;
-        }
-        control_frames.clear();
-    };
+    bool vae_decode_only = true;
+    SDImageOwner init_image({0, 0, 3, nullptr});
+    SDImageOwner end_image({0, 0, 3, nullptr});
+    SDImageOwner control_image({0, 0, 3, nullptr});
+    SDImageOwner mask_image({0, 0, 1, nullptr});
+    SDImageVec ref_images;
+    SDImageVec pmid_images;
+    SDImageVec control_frames;
 
     auto load_image_and_update_size = [&](const std::string& path,
-                                          sd_image_t& image,
+                                          SDImageOwner& image,
                                           bool resize_image    = true,
                                           int expected_channel = 3) -> bool {
         int expected_width  = 0;
@@ -527,13 +577,12 @@ int main(int argc, const char* argv[]) {
             expected_height = gen_params.height;
         }
 
-        if (!load_sd_image_from_file(&image, path.c_str(), expected_width, expected_height, expected_channel)) {
+        if (!load_sd_image_from_file(image.put(), path.c_str(), expected_width, expected_height, expected_channel)) {
             LOG_ERROR("load image from '%s' failed", path.c_str());
-            release_all_resources();
             return false;
         }
 
-        gen_params.set_width_and_height_if_unset(image.width, image.height);
+        gen_params.set_width_and_height_if_unset(image.get().width, image.get().height);
         return true;
     };
 
@@ -554,47 +603,46 @@ int main(int argc, const char* argv[]) {
     if (gen_params.ref_image_paths.size() > 0) {
         vae_decode_only = false;
         for (auto& path : gen_params.ref_image_paths) {
-            sd_image_t ref_image = {0, 0, 3, nullptr};
+            SDImageOwner ref_image({0, 0, 3, nullptr});
             if (!load_image_and_update_size(path, ref_image, false)) {
                 return 1;
             }
-            ref_images.push_back(ref_image);
+            ref_images.push_back(std::move(ref_image));
         }
     }
 
     if (gen_params.mask_image_path.size() > 0) {
-        if (!load_sd_image_from_file(&mask_image,
+        if (!load_sd_image_from_file(mask_image.put(),
                                      gen_params.mask_image_path.c_str(),
                                      gen_params.get_resolved_width(),
                                      gen_params.get_resolved_height(),
                                      1)) {
             LOG_ERROR("load image from '%s' failed", gen_params.mask_image_path.c_str());
-            release_all_resources();
             return 1;
         }
     } else {
-        mask_image.data = (uint8_t*)malloc(gen_params.get_resolved_width() * gen_params.get_resolved_height());
-        if (mask_image.data == nullptr) {
+        sd_image_t generated_mask = {0, 0, 1, nullptr};
+        generated_mask.data       = (uint8_t*)malloc(gen_params.get_resolved_width() * gen_params.get_resolved_height());
+        if (generated_mask.data == nullptr) {
             LOG_ERROR("malloc mask image failed");
-            release_all_resources();
             return 1;
         }
-        mask_image.width  = gen_params.get_resolved_width();
-        mask_image.height = gen_params.get_resolved_height();
-        memset(mask_image.data, 255, gen_params.get_resolved_width() * gen_params.get_resolved_height());
+        generated_mask.width  = gen_params.get_resolved_width();
+        generated_mask.height = gen_params.get_resolved_height();
+        memset(generated_mask.data, 255, gen_params.get_resolved_width() * gen_params.get_resolved_height());
+        mask_image.reset(generated_mask);
     }
 
     if (gen_params.control_image_path.size() > 0) {
-        if (!load_sd_image_from_file(&control_image,
+        if (!load_sd_image_from_file(control_image.put(),
                                      gen_params.control_image_path.c_str(),
                                      gen_params.get_resolved_width(),
                                      gen_params.get_resolved_height())) {
             LOG_ERROR("load image from '%s' failed", gen_params.control_image_path.c_str());
-            release_all_resources();
             return 1;
         }
         if (cli_params.canny_preprocess) {  // apply preprocessor
-            preprocess_canny(control_image,
+            preprocess_canny(control_image.get(),
                              0.08f,
                              0.08f,
                              0.8f,
@@ -610,7 +658,6 @@ int main(int argc, const char* argv[]) {
                                   gen_params.get_resolved_height(),
                                   gen_params.video_frames,
                                   cli_params.verbose)) {
-            release_all_resources();
             return 1;
         }
     }
@@ -622,7 +669,6 @@ int main(int argc, const char* argv[]) {
                                   0,
                                   0,
                                   cli_params.verbose)) {
-            release_all_resources();
             return 1;
         }
     }
@@ -633,39 +679,30 @@ int main(int argc, const char* argv[]) {
 
     sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(vae_decode_only, true, cli_params.taesd_preview);
 
-    sd_image_t* results = nullptr;
-    int num_results     = 0;
+    SDImageVec results;
+    int num_results = 0;
 
     if (cli_params.mode == UPSCALE) {
         num_results = 1;
-        results     = (sd_image_t*)calloc(num_results, sizeof(sd_image_t));
-        if (results == nullptr) {
-            LOG_INFO("failed to allocate results array");
-            release_all_resources();
-            return 1;
-        }
-
-        results[0]      = init_image;
-        init_image.data = nullptr;
+        results.push_back(init_image.release());
     } else {
-        sd_ctx_t* sd_ctx = new_sd_ctx(&sd_ctx_params);
+        SDCtxPtr sd_ctx(new_sd_ctx(&sd_ctx_params));
 
         if (sd_ctx == nullptr) {
             LOG_INFO("new_sd_ctx_t failed");
-            release_all_resources();
             return 1;
         }
 
         if (gen_params.sample_params.sample_method == SAMPLE_METHOD_COUNT) {
-            gen_params.sample_params.sample_method = sd_get_default_sample_method(sd_ctx);
+            gen_params.sample_params.sample_method = sd_get_default_sample_method(sd_ctx.get());
         }
 
         if (gen_params.high_noise_sample_params.sample_method == SAMPLE_METHOD_COUNT) {
-            gen_params.high_noise_sample_params.sample_method = sd_get_default_sample_method(sd_ctx);
+            gen_params.high_noise_sample_params.sample_method = sd_get_default_sample_method(sd_ctx.get());
         }
 
         if (gen_params.sample_params.scheduler == SCHEDULER_COUNT) {
-            gen_params.sample_params.scheduler = sd_get_default_scheduler(sd_ctx, gen_params.sample_params.sample_method);
+            gen_params.sample_params.scheduler = sd_get_default_scheduler(sd_ctx.get(), gen_params.sample_params.sample_method);
         }
 
         if (cli_params.mode == IMG_GEN) {
@@ -675,19 +712,19 @@ int main(int argc, const char* argv[]) {
                 gen_params.prompt.c_str(),
                 gen_params.negative_prompt.c_str(),
                 gen_params.clip_skip,
-                init_image,
+                init_image.get(),
                 ref_images.data(),
                 (int)ref_images.size(),
                 gen_params.auto_resize_ref_image,
                 gen_params.increase_ref_index,
-                mask_image,
+                mask_image.get(),
                 gen_params.get_resolved_width(),
                 gen_params.get_resolved_height(),
                 gen_params.sample_params,
                 gen_params.strength,
                 gen_params.seed,
                 gen_params.batch_count,
-                control_image,
+                control_image.get(),
                 gen_params.control_strength,
                 {
                     pmid_images.data(),
@@ -699,8 +736,8 @@ int main(int argc, const char* argv[]) {
                 gen_params.cache_params,
             };
 
-            results     = generate_image(sd_ctx, &img_gen_params);
             num_results = gen_params.batch_count;
+            results.adopt(generate_image(sd_ctx.get(), &img_gen_params), num_results);
         } else if (cli_params.mode == VID_GEN) {
             sd_vid_gen_params_t vid_gen_params = {
                 gen_params.lora_vec.data(),
@@ -708,8 +745,8 @@ int main(int argc, const char* argv[]) {
                 gen_params.prompt.c_str(),
                 gen_params.negative_prompt.c_str(),
                 gen_params.clip_skip,
-                init_image,
-                end_image,
+                init_image.get(),
+                end_image.get(),
                 control_frames.data(),
                 (int)control_frames.size(),
                 gen_params.get_resolved_width(),
@@ -725,25 +762,23 @@ int main(int argc, const char* argv[]) {
                 gen_params.cache_params,
             };
 
-            results = generate_video(sd_ctx, &vid_gen_params, &num_results);
+            sd_image_t* generated_video = generate_video(sd_ctx.get(), &vid_gen_params, &num_results);
+            results.adopt(generated_video, num_results);
         }
 
-        if (results == nullptr) {
+        if (!results) {
             LOG_ERROR("generate failed");
-            free_sd_ctx(sd_ctx);
             return 1;
         }
-
-        free_sd_ctx(sd_ctx);
     }
 
     int upscale_factor = 4;  // unused for RealESRGAN_x4plus_anime_6B.pth
     if (ctx_params.esrgan_path.size() > 0 && gen_params.upscale_repeats > 0) {
-        upscaler_ctx_t* upscaler_ctx = new_upscaler_ctx(ctx_params.esrgan_path.c_str(),
-                                                        ctx_params.offload_params_to_cpu,
-                                                        ctx_params.diffusion_conv_direct,
-                                                        ctx_params.n_threads,
-                                                        gen_params.upscale_tile_size);
+        UpscalerCtxPtr upscaler_ctx(new_upscaler_ctx(ctx_params.esrgan_path.c_str(),
+                                                     ctx_params.offload_params_to_cpu,
+                                                     ctx_params.diffusion_conv_direct,
+                                                     ctx_params.n_threads,
+                                                     gen_params.upscale_tile_size));
 
         if (upscaler_ctx == nullptr) {
             LOG_ERROR("new_upscaler_ctx failed");
@@ -752,32 +787,24 @@ int main(int argc, const char* argv[]) {
                 if (results[i].data == nullptr) {
                     continue;
                 }
-                sd_image_t current_image = results[i];
+                SDImageOwner current_image(results[i]);
+                results[i] = {0, 0, 0, nullptr};
                 for (int u = 0; u < gen_params.upscale_repeats; ++u) {
-                    sd_image_t upscaled_image = upscale(upscaler_ctx, current_image, upscale_factor);
-                    if (upscaled_image.data == nullptr) {
+                    SDImageOwner upscaled_image(upscale(upscaler_ctx.get(), current_image.get(), upscale_factor));
+                    if (upscaled_image.get().data == nullptr) {
                         LOG_ERROR("upscale failed");
                         break;
                     }
-                    free(current_image.data);
-                    current_image = upscaled_image;
+                    current_image = std::move(upscaled_image);
                 }
-                results[i] = current_image;  // Set the final upscaled image as the result
+                results[i] = current_image.release();  // Set the final upscaled image as the result
             }
         }
     }
 
-    if (!save_results(cli_params, ctx_params, gen_params, results, num_results)) {
+    if (!save_results(cli_params, ctx_params, gen_params, results.data(), num_results)) {
         return 1;
     }
 
-    for (int i = 0; i < num_results; i++) {
-        free(results[i].data);
-        results[i].data = nullptr;
-    }
-    free(results);
-
-    release_all_resources();
-
     return 0;
 }

examples/common/common.h

@@ -0,0 +1,207 @@
+#ifndef __EXAMPLES_COMMON_COMMON_H__
+#define __EXAMPLES_COMMON_COMMON_H__
+
+#include <cmath>
+#include <cstdint>
+#include <functional>
+#include <map>
+#include <string>
+#include <vector>
+
+#include "log.h"
+#include "stable-diffusion.h"
+
+#define SAFE_STR(s) ((s) ? (s) : "")
+#define BOOL_STR(b) ((b) ? "true" : "false")
+
+extern const char* const modes_str[];
+#define SD_ALL_MODES_STR "img_gen, vid_gen, convert, upscale, metadata"
+
+enum SDMode {
+    IMG_GEN,
+    VID_GEN,
+    CONVERT,
+    UPSCALE,
+    METADATA,
+    MODE_COUNT
+};
+
+struct StringOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    std::string* target;
+};
+
+struct IntOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    int* target;
+};
+
+struct FloatOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    float* target;
+};
+
+struct BoolOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    bool keep_true;
+    bool* target;
+};
+
+struct ManualOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    std::function<int(int argc, const char** argv, int index)> cb;
+};
+
+struct ArgOptions {
+    std::vector<StringOption> string_options;
+    std::vector<IntOption> int_options;
+    std::vector<FloatOption> float_options;
+    std::vector<BoolOption> bool_options;
+    std::vector<ManualOption> manual_options;
+
+    static std::string wrap_text(const std::string& text, size_t width, size_t indent);
+    void print() const;
+};
+
+bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& options_list);
+
+struct SDContextParams {
+    int n_threads = -1;
+    std::string model_path;
+    std::string clip_l_path;
+    std::string clip_g_path;
+    std::string clip_vision_path;
+    std::string t5xxl_path;
+    std::string llm_path;
+    std::string llm_vision_path;
+    std::string diffusion_model_path;
+    std::string high_noise_diffusion_model_path;
+    std::string vae_path;
+    std::string taesd_path;
+    std::string esrgan_path;
+    std::string control_net_path;
+    std::string embedding_dir;
+    std::string photo_maker_path;
+    sd_type_t wtype = SD_TYPE_COUNT;
+    std::string tensor_type_rules;
+    std::string lora_model_dir = ".";
+
+    std::map<std::string, std::string> embedding_map;
+    std::vector<sd_embedding_t> embedding_vec;
+
+    rng_type_t rng_type         = CUDA_RNG;
+    rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
+    bool offload_params_to_cpu  = false;
+    bool enable_mmap            = false;
+    bool control_net_cpu        = false;
+    bool clip_on_cpu            = false;
+    bool vae_on_cpu             = false;
+    bool flash_attn             = false;
+    bool diffusion_flash_attn   = false;
+    bool diffusion_conv_direct  = false;
+    bool vae_conv_direct        = false;
+
+    bool circular   = false;
+    bool circular_x = false;
+    bool circular_y = false;
+
+    bool chroma_use_dit_mask = true;
+    bool chroma_use_t5_mask  = false;
+    int chroma_t5_mask_pad   = 1;
+
+    bool qwen_image_zero_cond_t = false;
+
+    prediction_t prediction           = PREDICTION_COUNT;
+    lora_apply_mode_t lora_apply_mode = LORA_APPLY_AUTO;
+
+    bool force_sdxl_vae_conv_scale = false;
+
+    float flow_shift = INFINITY;
+    ArgOptions get_options();
+    void build_embedding_map();
+    bool process_and_check(SDMode mode);
+    std::string to_string() const;
+    sd_ctx_params_t to_sd_ctx_params_t(bool vae_decode_only, bool free_params_immediately, bool taesd_preview);
+};
+
+struct SDGenerationParams {
+    std::string prompt;
+    std::string prompt_with_lora;  // for metadata record only
+    std::string negative_prompt;
+    int clip_skip   = -1;  // <= 0 represents unspecified
+    int width       = -1;
+    int height      = -1;
+    int batch_count = 1;
+    std::string init_image_path;
+    std::string end_image_path;
+    std::string mask_image_path;
+    std::string control_image_path;
+    std::vector<std::string> ref_image_paths;
+    std::string control_video_path;
+    bool auto_resize_ref_image = true;
+    bool increase_ref_index    = false;
+    bool embed_image_metadata  = true;
+
+    std::vector<int> skip_layers = {7, 8, 9};
+    sd_sample_params_t sample_params;
+
+    std::vector<int> high_noise_skip_layers = {7, 8, 9};
+    sd_sample_params_t high_noise_sample_params;
+
+    std::vector<float> custom_sigmas;
+
+    std::string cache_mode;
+    std::string cache_option;
+    std::string scm_mask;
+    bool scm_policy_dynamic = true;
+    sd_cache_params_t cache_params{};
+
+    float moe_boundary  = 0.875f;
+    int video_frames    = 1;
+    int fps             = 16;
+    float vace_strength = 1.f;
+
+    float strength         = 0.75f;
+    float control_strength = 0.9f;
+
+    int64_t seed = 42;
+
+    sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
+
+    // Photo Maker
+    std::string pm_id_images_dir;
+    std::string pm_id_embed_path;
+    float pm_style_strength = 20.f;
+
+    int upscale_repeats   = 1;
+    int upscale_tile_size = 128;
+
+    std::map<std::string, float> lora_map;
+    std::map<std::string, float> high_noise_lora_map;
+    std::vector<sd_lora_t> lora_vec;
+    SDGenerationParams();
+    ArgOptions get_options();
+    bool from_json_str(const std::string& json_str);
+    void extract_and_remove_lora(const std::string& lora_model_dir);
+    bool width_and_height_are_set() const;
+    void set_width_and_height_if_unset(int w, int h);
+    int get_resolved_width() const;
+    int get_resolved_height() const;
+    bool process_and_check(SDMode mode, const std::string& lora_model_dir);
+    std::string to_string() const;
+};
+
+std::string version_string();
+std::string get_image_params(const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed);
+
+#endif  // __EXAMPLES_COMMON_COMMON_H__

examples/common/log.cpp

@@ -0,0 +1,115 @@
+#include "log.h"
+
+#include <vector>
+
+bool log_verbose = false;
+bool log_color   = false;
+
+std::string sd_basename(const std::string& path) {
+    size_t pos = path.find_last_of('/');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    pos = path.find_last_of('\\');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    return path;
+}
+
+void print_utf8(FILE* stream, const char* utf8) {
+    if (!utf8) {
+        return;
+    }
+
+#ifdef _WIN32
+    HANDLE h = (stream == stderr)
+                   ? GetStdHandle(STD_ERROR_HANDLE)
+                   : GetStdHandle(STD_OUTPUT_HANDLE);
+
+    DWORD mode;
+    BOOL is_console = GetConsoleMode(h, &mode);
+
+    if (is_console) {
+        int wlen = MultiByteToWideChar(CP_UTF8, 0, utf8, -1, NULL, 0);
+        if (wlen <= 0) {
+            return;
+        }
+
+        std::vector<wchar_t> wbuf(static_cast<size_t>(wlen));
+
+        MultiByteToWideChar(CP_UTF8, 0, utf8, -1, wbuf.data(), wlen);
+
+        DWORD written;
+        WriteConsoleW(h, wbuf.data(), wlen - 1, &written, NULL);
+    } else {
+        DWORD written;
+        WriteFile(h, utf8, (DWORD)strlen(utf8), &written, NULL);
+    }
+#else
+    fputs(utf8, stream);
+#endif
+}
+
+void log_print(enum sd_log_level_t level, const char* log, bool verbose, bool color) {
+    int tag_color;
+    const char* level_str;
+    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
+
+    if (!log || (!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:
+            tag_color = 33;
+            level_str = "?????";
+            break;
+    }
+
+    if (color) {
+        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
+    } else {
+        fprintf(out_stream, "[%-5s] ", level_str);
+    }
+    print_utf8(out_stream, log);
+    fflush(out_stream);
+}
+
+void example_log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
+    constexpr size_t LOG_BUFFER_SIZE = 4096;
+
+    va_list args;
+    va_start(args, format);
+
+    static char log_buffer[LOG_BUFFER_SIZE + 1];
+    int written = snprintf(log_buffer, LOG_BUFFER_SIZE, "%s:%-4d - ", sd_basename(file).c_str(), line);
+
+    if (written >= 0 && written < static_cast<int>(LOG_BUFFER_SIZE)) {
+        vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
+    }
+    size_t len = strlen(log_buffer);
+    if (len == 0 || log_buffer[len - 1] != '\n') {
+        strncat(log_buffer, "\n", LOG_BUFFER_SIZE - len);
+    }
+
+    log_print(level, log_buffer, log_verbose, log_color);
+
+    va_end(args);
+}

examples/common/log.h

@@ -0,0 +1,32 @@
+#ifndef __EXAMPLE_LOG_H__
+#define __EXAMPLE_LOG_H__
+
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <string>
+
+#if defined(_WIN32)
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#endif  // _WIN32
+
+#include "stable-diffusion.h"
+
+extern bool log_verbose;
+extern bool log_color;
+
+std::string sd_basename(const std::string& path);
+void print_utf8(FILE* stream, const char* utf8);
+void log_print(sd_log_level_t level, const char* log, bool verbose, bool color);
+void example_log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);
+
+#define LOG_DEBUG(format, ...) example_log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_INFO(format, ...) example_log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_WARN(format, ...) example_log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_ERROR(format, ...) example_log_printf(SD_LOG_ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
+
+#endif  // __EXAMPLE_LOG_H__

examples/common/media_io.h

@@ -0,0 +1,84 @@
+#ifndef __MEDIA_IO_H__
+#define __MEDIA_IO_H__
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "stable-diffusion.h"
+
+enum class EncodedImageFormat {
+    JPEG,
+    PNG,
+    WEBP,
+    UNKNOWN,
+};
+
+EncodedImageFormat encoded_image_format_from_path(const std::string& path);
+
+std::vector<uint8_t> encode_image_to_vector(EncodedImageFormat format,
+                                            const uint8_t* image,
+                                            int width,
+                                            int height,
+                                            int channels,
+                                            const std::string& parameters = "",
+                                            int quality                   = 90);
+
+bool write_image_to_file(const std::string& path,
+                         const uint8_t* image,
+                         int width,
+                         int height,
+                         int channels,
+                         const std::string& parameters = "",
+                         int quality                   = 90);
+
+uint8_t* load_image_from_file(const char* image_path,
+                              int& width,
+                              int& height,
+                              int expected_width   = 0,
+                              int expected_height  = 0,
+                              int expected_channel = 3);
+
+bool load_sd_image_from_file(sd_image_t* image,
+                             const char* image_path,
+                             int expected_width   = 0,
+                             int expected_height  = 0,
+                             int expected_channel = 3);
+
+uint8_t* load_image_from_memory(const char* image_bytes,
+                                int len,
+                                int& width,
+                                int& height,
+                                int expected_width   = 0,
+                                int expected_height  = 0,
+                                int expected_channel = 3);
+
+int create_mjpg_avi_from_sd_images(const char* filename,
+                                   sd_image_t* images,
+                                   int num_images,
+                                   int fps,
+                                   int quality = 90);
+
+#ifdef SD_USE_WEBP
+int create_animated_webp_from_sd_images(const char* filename,
+                                        sd_image_t* images,
+                                        int num_images,
+                                        int fps,
+                                        int quality = 90);
+#endif
+
+#ifdef SD_USE_WEBM
+int create_webm_from_sd_images(const char* filename,
+                               sd_image_t* images,
+                               int num_images,
+                               int fps,
+                               int quality = 90);
+#endif
+
+int create_video_from_sd_images(const char* filename,
+                                sd_image_t* images,
+                                int num_images,
+                                int fps,
+                                int quality = 90);
+
+#endif  // __MEDIA_IO_H__

examples/common/resource_owners.hpp

@@ -0,0 +1,207 @@
+#ifndef __EXAMPLE_RESOURCE_OWNERS_H__
+#define __EXAMPLE_RESOURCE_OWNERS_H__
+
+#include <cstdio>
+#include <cstdlib>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "stable-diffusion.h"
+
+struct FreeDeleter {
+    void operator()(void* ptr) const {
+        free(ptr);
+    }
+};
+
+struct FileCloser {
+    void operator()(FILE* file) const {
+        if (file != nullptr) {
+            fclose(file);
+        }
+    }
+};
+
+struct SDCtxDeleter {
+    void operator()(sd_ctx_t* ctx) const {
+        if (ctx != nullptr) {
+            free_sd_ctx(ctx);
+        }
+    }
+};
+
+struct UpscalerCtxDeleter {
+    void operator()(upscaler_ctx_t* ctx) const {
+        if (ctx != nullptr) {
+            free_upscaler_ctx(ctx);
+        }
+    }
+};
+
+template <typename T>
+using FreeUniquePtr = std::unique_ptr<T, FreeDeleter>;
+
+using FilePtr        = std::unique_ptr<FILE, FileCloser>;
+using SDCtxPtr       = std::unique_ptr<sd_ctx_t, SDCtxDeleter>;
+using UpscalerCtxPtr = std::unique_ptr<upscaler_ctx_t, UpscalerCtxDeleter>;
+
+class SDImageOwner {
+public:
+    SDImageOwner() = default;
+    explicit SDImageOwner(sd_image_t image)
+        : image_(image) {
+    }
+
+    SDImageOwner(const SDImageOwner&)            = delete;
+    SDImageOwner& operator=(const SDImageOwner&) = delete;
+
+    SDImageOwner(SDImageOwner&& other) noexcept
+        : image_(other.release()) {
+    }
+
+    SDImageOwner& operator=(SDImageOwner&& other) noexcept {
+        if (this != &other) {
+            reset();
+            image_ = other.release();
+        }
+        return *this;
+    }
+
+    ~SDImageOwner() {
+        reset();
+    }
+
+    sd_image_t* put() {
+        if (image_.data != nullptr) {
+            free(image_.data);
+            image_.data = nullptr;
+        }
+        image_.width  = 0;
+        image_.height = 0;
+        return &image_;
+    }
+
+    sd_image_t& get() {
+        return image_;
+    }
+
+    const sd_image_t& get() const {
+        return image_;
+    }
+
+    sd_image_t release() {
+        sd_image_t image = image_;
+        image_           = {0, 0, 0, nullptr};
+        return image;
+    }
+
+    void reset(sd_image_t image = {0, 0, 0, nullptr}) {
+        if (image_.data != nullptr) {
+            free(image_.data);
+        }
+        image_ = image;
+    }
+
+private:
+    sd_image_t image_ = {0, 0, 0, nullptr};
+};
+
+class SDImageVec {
+public:
+    SDImageVec() = default;
+
+    SDImageVec(const SDImageVec&)            = delete;
+    SDImageVec& operator=(const SDImageVec&) = delete;
+
+    SDImageVec(SDImageVec&& other) noexcept
+        : images_(std::move(other.images_)) {
+    }
+
+    SDImageVec& operator=(SDImageVec&& other) noexcept {
+        if (this != &other) {
+            clear();
+            images_ = std::move(other.images_);
+        }
+        return *this;
+    }
+
+    ~SDImageVec() {
+        clear();
+    }
+
+    void push_back(sd_image_t image) {
+        images_.push_back(image);
+    }
+
+    void push_back(SDImageOwner&& image) {
+        images_.push_back(image.release());
+    }
+
+    void reserve(size_t count) {
+        images_.reserve(count);
+    }
+
+    void adopt(sd_image_t* images, int count) {
+        clear();
+        if (images == nullptr || count <= 0) {
+            free(images);
+            return;
+        }
+
+        images_.reserve(static_cast<size_t>(count));
+        for (int i = 0; i < count; ++i) {
+            images_.push_back(images[i]);
+        }
+        free(images);
+    }
+
+    size_t size() const {
+        return images_.size();
+    }
+
+    bool empty() const {
+        return images_.empty();
+    }
+
+    explicit operator bool() const {
+        return !images_.empty();
+    }
+
+    sd_image_t* data() {
+        return images_.data();
+    }
+
+    const sd_image_t* data() const {
+        return images_.data();
+    }
+
+    sd_image_t& operator[](size_t index) {
+        return images_[index];
+    }
+
+    const sd_image_t& operator[](size_t index) const {
+        return images_[index];
+    }
+
+    std::vector<sd_image_t>& raw() {
+        return images_;
+    }
+
+    const std::vector<sd_image_t>& raw() const {
+        return images_;
+    }
+
+    void clear() {
+        for (sd_image_t& image : images_) {
+            free(image.data);
+            image.data = nullptr;
+        }
+        images_.clear();
+    }
+
+private:
+    std::vector<sd_image_t> images_;
+};
+
+#endif  // __EXAMPLE_RESOURCE_OWNERS_H__

examples/server/CMakeLists.txt

@@ -56,7 +56,12 @@ else()
     message(STATUS "Frontend disabled or directory not found: ${FRONTEND_DIR}")
 endif()
 
-add_executable(${TARGET} main.cpp)
+add_executable(${TARGET}
+    ../common/common.cpp
+    ../common/log.cpp
+    ../common/media_io.cpp
+    main.cpp
+)
 
 if(HAVE_FRONTEND_BUILD)
     add_dependencies(${TARGET} ${TARGET}_frontend)
@@ -70,10 +75,18 @@ endif()
 
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
+if(SD_WEBP)
+    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBP)
+    target_link_libraries(${TARGET} PRIVATE webp libwebpmux)
+endif()
+if(SD_WEBM)
+    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBM)
+    target_link_libraries(${TARGET} PRIVATE webm)
+endif()
 
 # due to httplib; it contains a pragma for MSVC, but other things need explicit flags
 if(WIN32 AND NOT MSVC)
     target_link_libraries(${TARGET} PRIVATE ws2_32)
 endif()
 
-target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)
+target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)

examples/server/README.md

@@ -189,7 +189,7 @@ 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 and TCD (default: 0)
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)
   --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)
@@ -197,7 +197,7 @@ 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 and TCD (default: 0)
+  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)
   --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

examples/server/main.cpp

@@ -8,15 +8,19 @@
 #include <sstream>
 #include <vector>
 
+#include <json.hpp>
 #include "httplib.h"
 #include "stable-diffusion.h"
 
-#include "common/common.hpp"
+#include "common/common.h"
+#include "common/media_io.h"
+#include "common/resource_owners.hpp"
 
 #ifdef HAVE_INDEX_HTML
 #include "frontend/dist/gen_index_html.h"
 #endif
 
+using json   = nlohmann::json;
 namespace fs = std::filesystem;
 
 // ----------------------- helpers -----------------------
@@ -217,62 +221,6 @@ std::string extract_and_remove_sd_cpp_extra_args(std::string& text) {
     return extracted;
 }
 
-enum class ImageFormat { JPEG,
-                         PNG };
-
-static int stbi_ext_write_png_to_func(stbi_write_func* func, void* context, int x, int y, int comp, const void* data, int stride_bytes, const char* parameters) {
-    int len;
-    unsigned char* png = stbi_write_png_to_mem((const unsigned char*)data, stride_bytes, x, y, comp, &len, parameters);
-    if (png == NULL)
-        return 0;
-    func(context, png, len);
-    STBIW_FREE(png);
-    return 1;
-}
-
-std::vector<uint8_t> write_image_to_vector(
-    ImageFormat format,
-    const uint8_t* image,
-    int width,
-    int height,
-    int channels,
-    std::string params = "",
-    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_ext_write_png_to_func(c_func, &ctx, width, height, channels, image, width * channels, params.size() > 0 ? params.c_str() : nullptr);
-            break;
-        default:
-            throw std::runtime_error("invalid image format");
-    }
-
-    if (!result) {
-        throw std::runtime_error("write imgage to mem failed");
-    }
-
-    return buffer;
-}
-
 void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
     SDSvrParams* svr_params = (SDSvrParams*)data;
     log_print(level, log, svr_params->verbose, svr_params->color);
@@ -341,18 +289,6 @@ std::string get_lora_full_path(ServerRuntime& rt, const std::string& path) {
     return (it != rt.lora_cache->end()) ? it->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);
-}
-
 void register_index_endpoints(httplib::Server& svr, const SDSvrParams& svr_params, const std::string& index_html) {
     const std::string serve_html_path = svr_params.serve_html_path;
     svr.Get("/", [serve_html_path, index_html](const httplib::Request&, httplib::Response& res) {
@@ -416,9 +352,9 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
             std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(prompt);
 
-            if (output_format != "png" && output_format != "jpeg") {
+            if (output_format != "png" && output_format != "jpeg" && output_format != "webp") {
                 res.status = 400;
-                res.set_content(R"({"error":"invalid output_format, must be one of [png, jpeg]"})", "application/json");
+                res.set_content(R"({"error":"invalid output_format, must be one of [png, jpeg, webp]"})", "application/json");
                 return;
             }
             if (n <= 0)
@@ -460,10 +396,10 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
             LOG_DEBUG("%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;
+            SDImageOwner init_image({(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr});
+            SDImageOwner control_image({(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr});
+            SDImageOwner mask_image({(uint32_t)gen_params.width, (uint32_t)gen_params.height, 1, nullptr});
+            SDImageVec pmid_images;
 
             sd_img_gen_params_t img_gen_params = {
                 gen_params.lora_vec.data(),
@@ -471,19 +407,19 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 gen_params.prompt.c_str(),
                 gen_params.negative_prompt.c_str(),
                 gen_params.clip_skip,
-                init_image,
+                init_image.get(),
                 nullptr,
                 0,
                 gen_params.auto_resize_ref_image,
                 gen_params.increase_ref_index,
-                mask_image,
+                mask_image.get(),
                 gen_params.width,
                 gen_params.height,
                 gen_params.sample_params,
                 gen_params.strength,
                 gen_params.seed,
                 gen_params.batch_count,
-                control_image,
+                control_image.get(),
                 gen_params.control_strength,
                 {
                     pmid_images.data(),
@@ -495,13 +431,19 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 gen_params.cache_params,
             };
 
-            sd_image_t* results = nullptr;
-            int num_results     = 0;
+            SDImageVec results;
+            int num_results = 0;
 
             {
                 std::lock_guard<std::mutex> lock(*runtime->sd_ctx_mutex);
-                results     = generate_image(runtime->sd_ctx, &img_gen_params);
                 num_results = gen_params.batch_count;
+                results.adopt(generate_image(runtime->sd_ctx, &img_gen_params), num_results);
+            }
+
+            if (!results) {
+                res.status = 500;
+                res.set_content(R"({"error":"generate failed"})", "application/json");
+                return;
             }
 
             for (int i = 0; i < num_results; i++) {
@@ -511,13 +453,17 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 std::string params = gen_params.embed_image_metadata
                                          ? get_image_params(*runtime->ctx_params, gen_params, gen_params.seed + i)
                                          : "";
-                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,
-                                                         params,
-                                                         output_compression);
+                auto image_bytes   = encode_image_to_vector(output_format == "jpeg"
+                                                                ? EncodedImageFormat::JPEG
+                                                            : output_format == "webp"
+                                                                ? EncodedImageFormat::WEBP
+                                                                : EncodedImageFormat::PNG,
+                                                          results[i].data,
+                                                          results[i].width,
+                                                          results[i].height,
+                                                          results[i].channel,
+                                                          params,
+                                                          output_compression);
                 if (image_bytes.empty()) {
                     LOG_ERROR("write image to mem failed");
                     continue;
@@ -528,8 +474,6 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 item["b64_json"] = b64;
                 out["data"].push_back(item);
             }
-            free_results(results, num_results);
-
             res.set_content(out.dump(), "application/json");
             res.status = 200;
 
@@ -650,9 +594,9 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
             LOG_DEBUG("%s\n", gen_params.to_string().c_str());
 
-            sd_image_t init_image    = {0, 0, 3, nullptr};
-            sd_image_t control_image = {0, 0, 3, nullptr};
-            std::vector<sd_image_t> pmid_images;
+            SDImageOwner init_image({0, 0, 3, nullptr});
+            SDImageOwner control_image({0, 0, 3, nullptr});
+            SDImageVec pmid_images;
 
             auto get_resolved_width = [&gen_params, runtime]() -> int {
                 if (gen_params.width > 0)
@@ -669,7 +613,7 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 return 512;
             };
 
-            std::vector<sd_image_t> ref_images;
+            SDImageVec ref_images;
             ref_images.reserve(images_bytes.size());
             for (auto& bytes : images_bytes) {
                 int img_w;
@@ -685,12 +629,12 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                     continue;
                 }
 
-                sd_image_t img{(uint32_t)img_w, (uint32_t)img_h, 3, raw_pixels};
-                gen_params.set_width_and_height_if_unset(img.width, img.height);
-                ref_images.push_back(img);
+                SDImageOwner img({(uint32_t)img_w, (uint32_t)img_h, 3, raw_pixels});
+                gen_params.set_width_and_height_if_unset(img.get().width, img.get().height);
+                ref_images.push_back(std::move(img));
             }
 
-            sd_image_t mask_image = {0};
+            SDImageOwner mask_image({0, 0, 1, nullptr});
             if (!mask_bytes.empty()) {
                 int expected_width  = 0;
                 int expected_height = 0;
@@ -706,13 +650,10 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                     static_cast<int>(mask_bytes.size()),
                     mask_w, mask_h,
                     expected_width, expected_height, 1);
-                mask_image = {(uint32_t)mask_w, (uint32_t)mask_h, 1, mask_raw};
-                gen_params.set_width_and_height_if_unset(mask_image.width, mask_image.height);
+                mask_image.reset({(uint32_t)mask_w, (uint32_t)mask_h, 1, mask_raw});
+                gen_params.set_width_and_height_if_unset(mask_image.get().width, mask_image.get().height);
             } else {
-                mask_image.width   = get_resolved_width();
-                mask_image.height  = get_resolved_height();
-                mask_image.channel = 1;
-                mask_image.data    = nullptr;
+                mask_image.reset({(uint32_t)get_resolved_width(), (uint32_t)get_resolved_height(), 1, nullptr});
             }
 
             sd_img_gen_params_t img_gen_params = {
@@ -721,19 +662,19 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 gen_params.prompt.c_str(),
                 gen_params.negative_prompt.c_str(),
                 gen_params.clip_skip,
-                init_image,
+                init_image.get(),
                 ref_images.data(),
                 (int)ref_images.size(),
                 gen_params.auto_resize_ref_image,
                 gen_params.increase_ref_index,
-                mask_image,
+                mask_image.get(),
                 get_resolved_width(),
                 get_resolved_height(),
                 gen_params.sample_params,
                 gen_params.strength,
                 gen_params.seed,
                 gen_params.batch_count,
-                control_image,
+                control_image.get(),
                 gen_params.control_strength,
                 {
                     pmid_images.data(),
@@ -745,13 +686,19 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 gen_params.cache_params,
             };
 
-            sd_image_t* results = nullptr;
-            int num_results     = 0;
+            SDImageVec results;
+            int num_results = 0;
 
             {
                 std::lock_guard<std::mutex> lock(*runtime->sd_ctx_mutex);
-                results     = generate_image(runtime->sd_ctx, &img_gen_params);
                 num_results = gen_params.batch_count;
+                results.adopt(generate_image(runtime->sd_ctx, &img_gen_params), num_results);
+            }
+
+            if (!results) {
+                res.status = 500;
+                res.set_content(R"({"error":"generate failed"})", "application/json");
+                return;
             }
 
             json out;
@@ -765,32 +712,24 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 std::string params = gen_params.embed_image_metadata
                                          ? get_image_params(*runtime->ctx_params, gen_params, gen_params.seed + i)
                                          : "";
-                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,
-                                                         params,
-                                                         output_compression);
+                auto image_bytes   = encode_image_to_vector(output_format == "jpeg"
+                                                                ? EncodedImageFormat::JPEG
+                                                            : output_format == "webp"
+                                                                ? EncodedImageFormat::WEBP
+                                                                : EncodedImageFormat::PNG,
+                                                          results[i].data,
+                                                          results[i].width,
+                                                          results[i].height,
+                                                          results[i].channel,
+                                                          params,
+                                                          output_compression);
                 std::string b64 = base64_encode(image_bytes);
                 json item;
                 item["b64_json"] = b64;
                 out["data"].push_back(item);
             }
-            free_results(results, num_results);
-
             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;
@@ -939,12 +878,11 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
             LOG_DEBUG("%s\n", gen_params.to_string().c_str());
 
-            sd_image_t init_image    = {0, 0, 3, nullptr};
-            sd_image_t control_image = {0, 0, 3, nullptr};
-            sd_image_t mask_image    = {0, 0, 1, nullptr};
-            std::vector<uint8_t> mask_data;
-            std::vector<sd_image_t> pmid_images;
-            std::vector<sd_image_t> ref_images;
+            SDImageOwner init_image({0, 0, 3, nullptr});
+            SDImageOwner control_image({0, 0, 3, nullptr});
+            SDImageOwner mask_image({0, 0, 1, nullptr});
+            SDImageVec pmid_images;
+            SDImageVec ref_images;
 
             auto get_resolved_width = [&gen_params, runtime]() -> int {
                 if (gen_params.width > 0)
@@ -961,7 +899,7 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 return 512;
             };
 
-            auto decode_image = [&gen_params](sd_image_t& image, std::string encoded) -> bool {
+            auto decode_image = [&gen_params](SDImageOwner& image, std::string encoded) -> bool {
                 auto comma_pos = encoded.find(',');
                 if (comma_pos != std::string::npos) {
                     encoded = encoded.substr(comma_pos + 1);
@@ -980,10 +918,10 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                     uint8_t* raw_data = load_image_from_memory(
                         (const char*)img_data.data(), (int)img_data.size(),
                         img_w, img_h,
-                        expected_width, expected_height, image.channel);
+                        expected_width, expected_height, image.get().channel);
                     if (raw_data) {
-                        image = {(uint32_t)img_w, (uint32_t)img_h, image.channel, raw_data};
-                        gen_params.set_width_and_height_if_unset(image.width, image.height);
+                        image.reset({(uint32_t)img_w, (uint32_t)img_h, image.get().channel, raw_data});
+                        gen_params.set_width_and_height_if_unset(image.get().width, image.get().height);
                         return true;
                     }
                 }
@@ -1000,19 +938,21 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                     std::string encoded = j["mask"].get<std::string>();
                     decode_image(mask_image, encoded);
                     bool inpainting_mask_invert = j.value("inpainting_mask_invert", 0) != 0;
-                    if (inpainting_mask_invert && mask_image.data != nullptr) {
-                        for (uint32_t i = 0; i < mask_image.width * mask_image.height; i++) {
-                            mask_image.data[i] = 255 - mask_image.data[i];
+                    if (inpainting_mask_invert && mask_image.get().data != nullptr) {
+                        for (uint32_t i = 0; i < mask_image.get().width * mask_image.get().height; i++) {
+                            mask_image.get().data[i] = 255 - mask_image.get().data[i];
                         }
                     }
                 } else {
-                    int m_width        = get_resolved_width();
-                    int m_height       = get_resolved_height();
-                    mask_data          = std::vector<uint8_t>(m_width * m_height, 255);
-                    mask_image.width   = m_width;
-                    mask_image.height  = m_height;
-                    mask_image.channel = 1;
-                    mask_image.data    = mask_data.data();
+                    int m_width               = get_resolved_width();
+                    int m_height              = get_resolved_height();
+                    sd_image_t generated_mask = {(uint32_t)m_width, (uint32_t)m_height, 1, nullptr};
+                    generated_mask.data       = (uint8_t*)malloc(static_cast<size_t>(m_width) * static_cast<size_t>(m_height));
+                    if (generated_mask.data == nullptr) {
+                        return bad("failed to allocate default mask");
+                    }
+                    memset(generated_mask.data, 255, static_cast<size_t>(m_width) * static_cast<size_t>(m_height));
+                    mask_image.reset(generated_mask);
                 }
 
                 float denoising_strength = j.value("denoising_strength", -1.f);
@@ -1024,10 +964,10 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
             if (j.contains("extra_images") && j["extra_images"].is_array()) {
                 for (auto extra_image : j["extra_images"]) {
-                    std::string encoded  = extra_image.get<std::string>();
-                    sd_image_t tmp_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+                    std::string encoded = extra_image.get<std::string>();
+                    SDImageOwner tmp_image({(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr});
                     if (decode_image(tmp_image, encoded)) {
-                        ref_images.push_back(tmp_image);
+                        ref_images.push_back(std::move(tmp_image));
                     }
                 }
             }
@@ -1038,19 +978,19 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 gen_params.prompt.c_str(),
                 gen_params.negative_prompt.c_str(),
                 gen_params.clip_skip,
-                init_image,
+                init_image.get(),
                 ref_images.data(),
                 (int)ref_images.size(),
                 gen_params.auto_resize_ref_image,
                 gen_params.increase_ref_index,
-                mask_image,
+                mask_image.get(),
                 get_resolved_width(),
                 get_resolved_height(),
                 gen_params.sample_params,
                 gen_params.strength,
                 gen_params.seed,
                 gen_params.batch_count,
-                control_image,
+                control_image.get(),
                 gen_params.control_strength,
                 {
                     pmid_images.data(),
@@ -1062,13 +1002,19 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 gen_params.cache_params,
             };
 
-            sd_image_t* results = nullptr;
-            int num_results     = 0;
+            SDImageVec results;
+            int num_results = 0;
 
             {
                 std::lock_guard<std::mutex> lock(*runtime->sd_ctx_mutex);
-                results     = generate_image(runtime->sd_ctx, &img_gen_params);
                 num_results = gen_params.batch_count;
+                results.adopt(generate_image(runtime->sd_ctx, &img_gen_params), num_results);
+            }
+
+            if (!results) {
+                res.status = 500;
+                res.set_content(R"({"error":"generate failed"})", "application/json");
+                return;
             }
 
             json out;
@@ -1084,12 +1030,12 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 std::string params = gen_params.embed_image_metadata
                                          ? get_image_params(*runtime->ctx_params, gen_params, gen_params.seed + i)
                                          : "";
-                auto image_bytes   = write_image_to_vector(ImageFormat::PNG,
-                                                           results[i].data,
-                                                           results[i].width,
-                                                           results[i].height,
-                                                           results[i].channel,
-                                                           params);
+                auto image_bytes   = encode_image_to_vector(EncodedImageFormat::PNG,
+                                                            results[i].data,
+                                                            results[i].width,
+                                                            results[i].height,
+                                                            results[i].channel,
+                                                            params);
 
                 if (image_bytes.empty()) {
                     LOG_ERROR("write image to mem failed");
@@ -1099,21 +1045,9 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 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;
 
-            if (init_image.data) {
-                stbi_image_free(init_image.data);
-            }
-            if (mask_image.data && mask_data.empty()) {
-                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;
@@ -1225,7 +1159,7 @@ int main(int argc, const char** argv) {
     LOG_DEBUG("%s", 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);
+    SDCtxPtr sd_ctx(new_sd_ctx(&sd_ctx_params));
 
     if (sd_ctx == nullptr) {
         LOG_ERROR("new_sd_ctx_t failed");
@@ -1237,7 +1171,7 @@ int main(int argc, const char** argv) {
     std::vector<LoraEntry> lora_cache;
     std::mutex lora_mutex;
     ServerRuntime runtime = {
-        sd_ctx,
+        sd_ctx.get(),
         &sd_ctx_mutex,
         &svr_params,
         &ctx_params,
@@ -1278,6 +1212,5 @@ int main(int argc, const char** argv) {
     LOG_INFO("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);
 
-    free_sd_ctx(sd_ctx);
     return 0;
 }

format-code.sh

@@ -1,4 +1,6 @@
-for f in src/*.cpp src/*.h src/*.hpp src/vocab/*.h src/vocab/*.cpp examples/cli/*.cpp examples/common/*.hpp examples/cli/*.h examples/server/*.cpp; do
+for f in src/*.cpp src/*.h src/*.hpp src/vocab/*.h src/vocab/*.cpp \
+         examples/cli/*.cpp examples/cli/*.h examples/server/*.cpp \
+         examples/common/*.hpp examples/common/*.h examples/common/*.cpp; do
   [[ "$f" == vocab* ]] && continue
   echo "formatting '$f'"
   # if [ "$f" != "stable-diffusion.h" ]; then

src/denoiser.hpp

@@ -668,8 +668,8 @@ struct DiscreteFlowDenoiser : public Denoiser {
     }
 
     void set_parameters() {
-        for (int i = 1; i < TIMESTEPS + 1; i++) {
-            sigmas[i - 1] = t_to_sigma(static_cast<float>(i));
+        for (int i = 0; i < TIMESTEPS; i++) {
+            sigmas[i] = t_to_sigma(static_cast<float>(i));
         }
     }
 
@@ -786,10 +786,43 @@ static std::pair<float, float> get_ancestral_step(float sigma_from,
     return {sigma_down, sigma_up};
 }
 
+static std::tuple<float, float, float> get_ancestral_step_flow(float sigma_from,
+                                                               float sigma_to,
+                                                               float eta = 1.0f) {
+    float sigma_down  = sigma_to;
+    float sigma_up    = 0.0f;
+    float alpha_scale = 1.0f;
+
+    if (eta <= 0.0f || sigma_from <= 0.0f || sigma_to <= 0.0f) {
+        return {sigma_down, sigma_up, alpha_scale};
+    }
+
+    // Flow Euler ancestral sampling becomes numerically unstable for eta > 1, so
+    // clamp to the valid maximum-noise regime instead of letting NaNs propagate.
+    eta = std::min(eta, 1.0f);
+
+    float sigma_ratio = sigma_to / sigma_from;
+    sigma_down        = sigma_to * (1.0f + (sigma_ratio - 1.0f) * eta);
+    sigma_down        = std::max(0.0f, std::min(sigma_to, sigma_down));
+
+    float denom = 1.0f - sigma_down;
+    if (denom <= 0.0f) {
+        return {sigma_to, sigma_up, alpha_scale};
+    }
+
+    alpha_scale = (1.0f - sigma_to) / denom;
+
+    float term = (sigma_down / sigma_to) * alpha_scale;
+    term       = std::max(-1.0f, std::min(1.0f, term));
+    sigma_up   = sigma_to * std::sqrt(std::max(1.0f - term * term, 0.0f));
+    return {sigma_down, sigma_up, alpha_scale};
+}
+
 static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
                                                 sd::Tensor<float> x,
                                                 const std::vector<float>& sigmas,
-                                                std::shared_ptr<RNG> rng) {
+                                                std::shared_ptr<RNG> rng,
+                                                float eta) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma       = sigmas[i];
@@ -799,7 +832,7 @@ static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
         }
         sd::Tensor<float> denoised  = std::move(denoised_opt);
         sd::Tensor<float> d         = (x - denoised) / sigma;
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1]);
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
         x += d * (sigma_down - sigmas[i]);
         if (sigmas[i + 1] > 0) {
             x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
@@ -808,6 +841,30 @@ static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
     return x;
 }
 
+static sd::Tensor<float> sample_euler_flow(denoise_cb_t model,
+                                           sd::Tensor<float> x,
+                                           const std::vector<float>& sigmas,
+                                           std::shared_ptr<RNG> rng,
+                                           float eta) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma       = sigmas[i];
+        auto denoised_opt = model(x, sigma, i + 1);
+        if (denoised_opt.empty()) {
+            return {};
+        }
+        sd::Tensor<float> denoised               = std::move(denoised_opt);
+        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step_flow(sigma, sigmas[i + 1], eta);
+        float sigma_ratio                        = sigma_down / sigma;
+        x                                        = sigma_ratio * x + (1.0f - sigma_ratio) * denoised;
+
+        if (sigma_up > 0.0f) {
+            x = alpha_scale * x + sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+        }
+    }
+    return x;
+}
+
 static sd::Tensor<float> sample_euler(denoise_cb_t model,
                                       sd::Tensor<float> x,
                                       const std::vector<float>& sigmas) {
@@ -885,7 +942,8 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
 static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
                                                    sd::Tensor<float> x,
                                                    const std::vector<float>& sigmas,
-                                                   std::shared_ptr<RNG> rng) {
+                                                   std::shared_ptr<RNG> rng,
+                                                   float eta) {
     auto t_fn     = [](float sigma) -> float { return -log(sigma); };
     auto sigma_fn = [](float t) -> float { return exp(-t); };
 
@@ -896,7 +954,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
             return {};
         }
         sd::Tensor<float> denoised  = std::move(denoised_opt);
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1]);
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
 
         if (sigma_down == 0) {
             x = denoised;
@@ -1368,10 +1426,14 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
                                             sd::Tensor<float> x,
                                             std::vector<float> sigmas,
                                             std::shared_ptr<RNG> rng,
-                                            float eta) {
+                                            float eta,
+                                            bool is_flow_denoiser) {
     switch (method) {
         case EULER_A_SAMPLE_METHOD:
-            return sample_euler_ancestral(model, std::move(x), sigmas, rng);
+            if (is_flow_denoiser)
+                return sample_euler_flow(model, std::move(x), sigmas, rng, eta);
+            else
+                return sample_euler_ancestral(model, std::move(x), sigmas, rng, eta);
         case EULER_SAMPLE_METHOD:
             return sample_euler(model, std::move(x), sigmas);
         case HEUN_SAMPLE_METHOD:
@@ -1379,7 +1441,7 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
         case DPM2_SAMPLE_METHOD:
             return sample_dpm2(model, std::move(x), sigmas);
         case DPMPP2S_A_SAMPLE_METHOD:
-            return sample_dpmpp_2s_ancestral(model, std::move(x), sigmas, rng);
+            return sample_dpmpp_2s_ancestral(model, std::move(x), sigmas, rng, eta);
         case DPMPP2M_SAMPLE_METHOD:
             return sample_dpmpp_2m(model, std::move(x), sigmas);
         case DPMPP2Mv2_SAMPLE_METHOD:

src/stable-diffusion.cpp

@@ -1593,6 +1593,7 @@ public:
                              float eta,
                              int shifted_timestep,
                              sample_method_t method,
+                             bool is_flow_denoiser,
                              const std::vector<float>& sigmas,
                              int start_merge_step,
                              const std::vector<sd::Tensor<float>>& ref_latents,
@@ -1791,7 +1792,7 @@ public:
             return denoised;
         };
 
-        auto x0_opt = sample_k_diffusion(method, denoise, x_t, sigmas, sampler_rng, eta);
+        auto x0_opt = sample_k_diffusion(method, denoise, x_t, sigmas, sampler_rng, eta, is_flow_denoiser);
         if (x0_opt.empty()) {
             LOG_ERROR("Diffusion model sampling failed");
             if (control_net) {
@@ -1909,6 +1910,11 @@ public:
             flow_denoiser->set_shift(flow_shift);
         }
     }
+
+    bool is_flow_denoiser() {
+        auto flow_denoiser = std::dynamic_pointer_cast<DiscreteFlowDenoiser>(denoiser);
+        return !!flow_denoiser;
+    }
 };
 
 /*================================================= SD API ==================================================*/
@@ -2225,6 +2231,7 @@ 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->eta                         = INFINITY;
     sample_params->custom_sigmas               = nullptr;
     sample_params->custom_sigmas_count         = 0;
     sample_params->flow_shift                  = INFINITY;
@@ -2438,6 +2445,26 @@ static scheduler_t resolve_scheduler(sd_ctx_t* sd_ctx,
     return scheduler;
 }
 
+static float resolve_eta(sd_ctx_t* sd_ctx,
+                         float eta,
+                         enum sample_method_t sample_method) {
+    if (eta == INFINITY) {
+        switch (sample_method) {
+            case DDIM_TRAILING_SAMPLE_METHOD:
+            case TCD_SAMPLE_METHOD:
+            case RES_MULTISTEP_SAMPLE_METHOD:
+            case RES_2S_SAMPLE_METHOD:
+                return 0.0f;
+            case EULER_A_SAMPLE_METHOD:
+            case DPMPP2S_A_SAMPLE_METHOD:
+                return 1.0f;
+            default:;
+        }
+        return 0.0f;
+    }
+    return eta;
+}
+
 struct GenerationRequest {
     std::string prompt;
     std::string negative_prompt;
@@ -2586,6 +2613,8 @@ struct GenerationRequest {
 struct SamplePlan {
     enum sample_method_t sample_method            = SAMPLE_METHOD_COUNT;
     enum sample_method_t high_noise_sample_method = SAMPLE_METHOD_COUNT;
+    float eta                                     = 0.f;
+    float high_noise_eta                          = 0.f;
     int sample_steps                              = 0;
     int high_noise_sample_steps                   = 0;
     int total_steps                               = 0;
@@ -2597,6 +2626,7 @@ struct SamplePlan {
                const sd_img_gen_params_t* sd_img_gen_params,
                const GenerationRequest& request) {
         sample_method = sd_img_gen_params->sample_params.sample_method;
+        eta           = sd_img_gen_params->sample_params.eta;
         sample_steps  = sd_img_gen_params->sample_params.sample_steps;
         resolve(sd_ctx, &request, &sd_img_gen_params->sample_params);
     }
@@ -2605,10 +2635,12 @@ struct SamplePlan {
                const sd_vid_gen_params_t* sd_vid_gen_params,
                const GenerationRequest& request) {
         sample_method = sd_vid_gen_params->sample_params.sample_method;
+        eta           = sd_vid_gen_params->sample_params.eta;
         sample_steps  = sd_vid_gen_params->sample_params.sample_steps;
         if (sd_ctx->sd->high_noise_diffusion_model) {
             high_noise_sample_steps  = sd_vid_gen_params->high_noise_sample_params.sample_steps;
             high_noise_sample_method = sd_vid_gen_params->high_noise_sample_params.sample_method;
+            high_noise_eta           = sd_vid_gen_params->high_noise_sample_params.eta;
         }
         moe_boundary = sd_vid_gen_params->moe_boundary;
         resolve(sd_ctx, &request, &sd_vid_gen_params->sample_params);
@@ -2644,6 +2676,8 @@ struct SamplePlan {
                                                                      sd_ctx->sd->version);
         }
 
+        eta = resolve_eta(sd_ctx, eta, sample_method);
+
         if (high_noise_sample_steps < 0) {
             for (size_t i = 0; i < sigmas.size(); ++i) {
                 if (sigmas[i] < moe_boundary) {
@@ -2658,6 +2692,7 @@ struct SamplePlan {
         if (high_noise_sample_steps > 0) {
             high_noise_sample_method = resolve_sample_method(sd_ctx,
                                                              high_noise_sample_method);
+            high_noise_eta           = resolve_eta(sd_ctx, high_noise_eta, high_noise_sample_method);
             LOG_INFO("sampling(high noise) using %s method", sampling_methods_str[high_noise_sample_method]);
         }
 
@@ -2811,7 +2846,8 @@ static std::optional<ImageGenerationLatents> prepare_image_generation_latents(sd
                                                          {request->width / request->vae_scale_factor,
                                                           request->height / request->vae_scale_factor,
                                                           1,
-                                                          1});
+                                                          1},
+                                                         sd::ops::InterpolateMode::NearestMax);
 
     sd::Tensor<float> init_latent;
     sd::Tensor<float> control_latent;
@@ -2956,8 +2992,12 @@ static std::optional<ImageGenerationLatents> prepare_image_generation_latents(sd
     latents.ref_latents          = std::move(ref_latents);
 
     if (sd_version_is_inpaint(sd_ctx->sd->version)) {
-        latents.denoise_mask = std::move(latent_mask);
+        latent_mask = sd::ops::max_pool_2d(latent_mask,
+                                           {3, 3},
+                                           {1, 1},
+                                           {1, 1});
     }
+    latents.denoise_mask = std::move(latent_mask);
 
     return latents;
 }
@@ -3123,9 +3163,10 @@ SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* s
                                                    latents.control_image,
                                                    request.control_strength,
                                                    request.guidance,
-                                                   request.eta,
+                                                   plan.eta,
                                                    request.shifted_timestep,
                                                    plan.sample_method,
+                                                   sd_ctx->sd->is_flow_denoiser(),
                                                    plan.sigmas,
                                                    plan.start_merge_step,
                                                    latents.ref_latents,
@@ -3482,9 +3523,10 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
                                                            sd::Tensor<float>(),
                                                            0.f,
                                                            request.high_noise_guidance,
-                                                           sd_vid_gen_params->high_noise_sample_params.eta,
+                                                           plan.high_noise_eta,
                                                            request.shifted_timestep,
                                                            plan.high_noise_sample_method,
+                                                           sd_ctx->sd->is_flow_denoiser(),
                                                            high_noise_sigmas,
                                                            -1,
                                                            std::vector<sd::Tensor<float>>{},
@@ -3523,9 +3565,10 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
                                                         sd::Tensor<float>(),
                                                         0.f,
                                                         sd_vid_gen_params->sample_params.guidance,
-                                                        sd_vid_gen_params->sample_params.eta,
+                                                        plan.eta,
                                                         sd_vid_gen_params->sample_params.shifted_timestep,
                                                         plan.sample_method,
+                                                        sd_ctx->sd->is_flow_denoiser(),
                                                         plan.sigmas,
                                                         -1,
                                                         std::vector<sd::Tensor<float>>{},

src/t5.hpp

@@ -1,4 +1,4 @@
-#ifndef __T5_HPP__
+#ifndef __T5_HPP__
 #define __T5_HPP__
 
 #include <cfloat>

src/tensor.hpp

@@ -815,6 +815,9 @@ namespace sd {
     namespace ops {
         enum class InterpolateMode {
             Nearest,
+            NearestMax,
+            NearestMin,
+            NearestAvg,
         };
 
         inline int64_t normalize_slice_bound(int64_t index, int64_t dim_size) {
@@ -1012,12 +1015,16 @@ namespace sd {
                                      std::vector<int64_t> output_shape,
                                      InterpolateMode mode = InterpolateMode::Nearest,
                                      bool align_corners   = false) {
-            if (mode != InterpolateMode::Nearest) {
-                tensor_throw_invalid_argument("Only nearest interpolate mode is implemented, got mode=" +
+            const bool is_nearest_like_mode = (mode == InterpolateMode::Nearest ||
+                                               mode == InterpolateMode::NearestMax ||
+                                               mode == InterpolateMode::NearestMin ||
+                                               mode == InterpolateMode::NearestAvg);
+            if (!is_nearest_like_mode) {
+                tensor_throw_invalid_argument("Only nearest-like interpolate modes are implemented, got mode=" +
                                               std::to_string(static_cast<int>(mode)));
             }
             if (align_corners) {
-                tensor_throw_invalid_argument("align_corners is not supported for nearest interpolate: input_shape=" +
+                tensor_throw_invalid_argument("align_corners is not supported for nearest-like interpolate: input_shape=" +
                                               tensor_shape_to_string(input.shape()) + ", output_shape=" +
                                               tensor_shape_to_string(output_shape));
             }
@@ -1044,14 +1051,102 @@ namespace sd {
                 }
             }
 
-            Tensor<T> output(std::move(output_shape));
-            for (int64_t flat = 0; flat < output.numel(); ++flat) {
-                std::vector<int64_t> output_coord = tensor_unravel_index(flat, output.shape());
-                std::vector<int64_t> input_coord(static_cast<size_t>(input.dim()), 0);
-                for (size_t i = 0; i < static_cast<size_t>(input.dim()); ++i) {
-                    input_coord[i] = output_coord[i] * input.shape()[i] / output.shape()[i];
+            bool has_downsampling = false;
+            for (int64_t i = 0; i < input.dim(); ++i) {
+                if (input.shape()[i] > output_shape[i]) {
+                    has_downsampling = true;
+                    break;
                 }
-                output[flat] = input.index(input_coord);
+            }
+
+            Tensor<T> output(std::move(output_shape));
+            if (mode == InterpolateMode::Nearest || !has_downsampling) {
+                for (int64_t flat = 0; flat < output.numel(); ++flat) {
+                    std::vector<int64_t> output_coord = tensor_unravel_index(flat, output.shape());
+                    std::vector<int64_t> input_coord(static_cast<size_t>(input.dim()), 0);
+                    for (size_t i = 0; i < static_cast<size_t>(input.dim()); ++i) {
+                        input_coord[i] = output_coord[i] * input.shape()[i] / output.shape()[i];
+                    }
+                    output[flat] = input.index(input_coord);
+                }
+
+                return output;
+            }
+
+            auto init_reduction = [&]() -> T {
+                switch (mode) {
+                    case InterpolateMode::NearestMax:
+                        return std::numeric_limits<T>::lowest();
+                    case InterpolateMode::NearestMin:
+                        return std::numeric_limits<T>::max();
+                    case InterpolateMode::NearestAvg:
+                        return T(0);
+                    case InterpolateMode::Nearest:
+                        return T(0);
+                }
+
+                tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
+                                              std::to_string(static_cast<int>(mode)));
+            };
+
+            auto reduce_value = [&](T& acc, const T& sample) {
+                switch (mode) {
+                    case InterpolateMode::NearestMax:
+                        acc = std::max(acc, sample);
+                        break;
+                    case InterpolateMode::NearestMin:
+                        acc = std::min(acc, sample);
+                        break;
+                    case InterpolateMode::NearestAvg:
+                        acc += sample;
+                        break;
+                    case InterpolateMode::Nearest:
+                        break;
+                }
+            };
+
+            // Reduction modes only differ from nearest mode when downsampling.
+            for (int64_t flat_out = 0; flat_out < output.numel(); ++flat_out) {
+                std::vector<int64_t> output_coord = tensor_unravel_index(flat_out, output.shape());
+
+                std::vector<int64_t> input_start(output.dim(), 0);
+                std::vector<int64_t> input_end(output.dim(), 0);
+
+                for (size_t i = 0; i < static_cast<size_t>(output.dim()); ++i) {
+                    const int64_t input_dim  = input.shape()[i];
+                    const int64_t output_dim = output.shape()[i];
+
+                    input_start[i] = std::max(int64_t(0), static_cast<int64_t>(output_coord[i] * input_dim / output_dim));
+                    input_end[i]   = std::min(input_dim, ((output_coord[i] + 1) * input_dim + output_dim - 1) / output_dim);
+                }
+
+                T value                               = init_reduction();
+                bool done_window                      = false;
+                std::vector<int64_t> current_in_coord = input_start;
+
+                while (!done_window) {
+                    reduce_value(value, input.index(current_in_coord));
+
+                    for (int d = static_cast<int>(output.dim()) - 1; d >= 0; --d) {
+                        if (++current_in_coord[d] < input_end[d]) {
+                            break;
+                        }
+                        current_in_coord[d] = input_start[d];
+                        if (d == 0) {
+                            done_window = true;
+                        }
+                    }
+                }
+
+                if (mode == InterpolateMode::NearestAvg) {
+                    int64_t window_size = 1;
+                    for (size_t i = 0; i < static_cast<size_t>(output.dim()); ++i) {
+                        window_size *= (input_end[i] - input_start[i]);
+                    }
+                    value /= static_cast<T>(window_size);
+                }
+
+                output[flat_out] = value;
             }
 
             return output;
@@ -1063,12 +1158,16 @@ namespace sd {
                                      const std::optional<std::vector<double>>& scale_factor,
                                      InterpolateMode mode = InterpolateMode::Nearest,
                                      bool align_corners   = false) {
-            if (mode != InterpolateMode::Nearest) {
-                tensor_throw_invalid_argument("Only nearest interpolate mode is implemented, got mode=" +
+            const bool is_nearest_like_mode = (mode == InterpolateMode::Nearest ||
+                                               mode == InterpolateMode::NearestMax ||
+                                               mode == InterpolateMode::NearestMin ||
+                                               mode == InterpolateMode::NearestAvg);
+            if (!is_nearest_like_mode) {
+                tensor_throw_invalid_argument("Only nearest-like interpolate modes are implemented, got mode=" +
                                               std::to_string(static_cast<int>(mode)));
             }
             if (align_corners) {
-                tensor_throw_invalid_argument("align_corners is not supported for nearest interpolate: input_shape=" +
+                tensor_throw_invalid_argument("align_corners is not supported for nearest-like interpolate: input_shape=" +
                                               tensor_shape_to_string(input.shape()));
             }
             if (size.has_value() == scale_factor.has_value()) {
@@ -1128,6 +1227,80 @@ namespace sd {
                                align_corners);
         }
 
+        template <typename T>
+        inline Tensor<T> max_pool_2d(const Tensor<T>& input,
+                                     std::vector<int64_t> kernel_size,
+                                     std::vector<int64_t> stride,
+                                     std::vector<int64_t> padding) {
+            if (input.dim() < 2) {
+                tensor_throw_invalid_argument("Tensor max_pool_2d requires input_dim >= 2: input_dim=" +
+                                              std::to_string(input.dim()) + ", input_shape=" +
+                                              tensor_shape_to_string(input.shape()));
+            }
+            if (kernel_size.size() != 2 || stride.size() != 2 || padding.size() != 2) {
+                tensor_throw_invalid_argument("Tensor max_pool_2d requires kernel_size, stride, and padding to have length 2");
+            }
+            for (size_t i = 0; i < 2; ++i) {
+                if (kernel_size[i] <= 0) {
+                    tensor_throw_invalid_argument("Tensor max_pool_2d kernel_size must be positive: kernel_size=" +
+                                                  tensor_shape_to_string(kernel_size));
+                }
+                if (stride[i] <= 0) {
+                    tensor_throw_invalid_argument("Tensor max_pool_2d stride must be positive: stride=" +
+                                                  tensor_shape_to_string(stride));
+                }
+                if (padding[i] < 0) {
+                    tensor_throw_invalid_argument("Tensor max_pool_2d padding must be non-negative: padding=" +
+                                                  tensor_shape_to_string(padding));
+                }
+            }
+
+            const int64_t in_height = input.shape()[0];
+            const int64_t in_width  = input.shape()[1];
+
+            const int64_t out_height = (in_height + 2 * padding[0] - kernel_size[0]) / stride[0] + 1;
+            const int64_t out_width  = (in_width + 2 * padding[1] - kernel_size[1]) / stride[1] + 1;
+
+            if (out_height <= 0 || out_width <= 0) {
+                tensor_throw_invalid_argument("max_pool_2d results in invalid output dimensions: " +
+                                              std::to_string(out_height) + "x" + std::to_string(out_width));
+            }
+
+            std::vector<int64_t> output_shape = input.shape();
+            output_shape[0]                   = out_height;
+            output_shape[1]                   = out_width;
+
+            Tensor<T> output(std::move(output_shape));
+
+            for (int64_t flat_out = 0; flat_out < output.numel(); ++flat_out) {
+                std::vector<int64_t> output_coord = tensor_unravel_index(flat_out, output.shape());
+                std::vector<int64_t> input_coord  = output_coord;
+
+                const int64_t oh = output_coord[0];
+                const int64_t ow = output_coord[1];
+
+                T max_val            = std::numeric_limits<T>::lowest();
+                bool has_valid_input = false;
+
+                for (int64_t kh = 0; kh < kernel_size[0]; ++kh) {
+                    for (int64_t kw = 0; kw < kernel_size[1]; ++kw) {
+                        const int64_t ih = oh * stride[0] + kh - padding[0];
+                        const int64_t iw = ow * stride[1] + kw - padding[1];
+
+                        if (ih >= 0 && ih < in_height && iw >= 0 && iw < in_width) {
+                            input_coord[0]  = ih;
+                            input_coord[1]  = iw;
+                            max_val         = std::max(max_val, input.index(input_coord));
+                            has_valid_input = true;
+                        }
+                    }
+                }
+
+                output[flat_out] = has_valid_input ? max_val : T(0);
+            }
+            return output;
+        }
+
         template <typename T>
         inline Tensor<T> concat(const Tensor<T>& lhs, const Tensor<T>& rhs, size_t dim) {
             if (lhs.dim() != rhs.dim()) {

src/tokenize_util.cpp

@@ -1,4 +1,4 @@
-#include <algorithm>
+#include <algorithm>
 #include <iostream>
 #include <string>
 #include <vector>

thirdparty/CMakeLists.txt

@@ -1,3 +1,45 @@
 set(Z_TARGET zip)
 add_library(${Z_TARGET} OBJECT zip.c zip.h miniz.h)
-target_include_directories(${Z_TARGET} PUBLIC .)
+target_include_directories(${Z_TARGET} PUBLIC .)
+
+if(SD_WEBP AND NOT SD_USE_SYSTEM_WEBP)
+    set(WEBP_BUILD_ANIM_UTILS OFF)
+    set(WEBP_BUILD_CWEBP OFF)
+    set(WEBP_BUILD_DWEBP OFF)
+    set(WEBP_BUILD_GIF2WEBP OFF)
+    set(WEBP_BUILD_IMG2WEBP OFF)
+    set(WEBP_BUILD_VWEBP OFF)
+    set(WEBP_BUILD_WEBPINFO OFF)
+    set(WEBP_BUILD_WEBPMUX OFF)
+    set(WEBP_BUILD_EXTRAS OFF)
+    set(WEBP_BUILD_WEBP_JS OFF)
+    set(WEBP_BUILD_FUZZTEST OFF)
+    set(WEBP_BUILD_LIBWEBPMUX ON)
+
+    add_subdirectory(libwebp EXCLUDE_FROM_ALL)
+endif()
+
+if(SD_WEBM AND NOT SD_USE_SYSTEM_WEBM)
+    if(MSVC)
+        set(MSVC_RUNTIME dll)
+    endif()
+    set(ENABLE_WEBMTS OFF)
+    set(ENABLE_WEBMINFO OFF)
+    set(ENABLE_TESTS OFF)
+    set(ENABLE_WEBM_PARSER OFF)
+    set(ENABLE_SAMPLE_PROGRAMS OFF)
+
+    set(SD_LIBWEBM_PARENT_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+
+    add_subdirectory(libwebm EXCLUDE_FROM_ALL)
+
+    # libwebm mutates the global CMAKE_CXX_FLAGS for non-MSVC compilers to force
+    # C++11. Restore the parent flags so the main project keeps its own C++17
+    # requirements, then pin the libwebm targets to C++17 explicitly.
+    set(CMAKE_CXX_FLAGS "${SD_LIBWEBM_PARENT_CXX_FLAGS}" CACHE STRING "" FORCE)
+    target_compile_features(mkvmuxer PRIVATE cxx_std_17)
+    target_compile_features(mkvparser PRIVATE cxx_std_17)
+    target_compile_features(webm PRIVATE cxx_std_17)
+
+    target_include_directories(webm INTERFACE ${CMAKE_CURRENT_SOURCE_DIR}/libwebm)
+endif()