feat: add ltx2.3 support (#1463)

* add GemmaTokenizer

* add basic ltx2.3 support

* change vocab file encoding

* fix ci

* fix ubuntu build

* add temporal tiling support

* add ltx audio support

* update ggml submodule url

* fix generate_video

* add i2v support

* minify bundled Gemma tokenizer vocab sources

* pass video fps into temporal rope embeddings

* fix av_ca_timestep_scale_multiplier

* add LTX2Scheduler support

* update docs

* fix ci

examples/cli/CMakeLists.txt

@@ -7,6 +7,10 @@ add_executable(${TARGET}
     image_metadata.cpp
     main.cpp
 )
+target_include_directories(${TARGET} PRIVATE
+    "${CMAKE_CURRENT_SOURCE_DIR}/.."
+    "${PROJECT_SOURCE_DIR}/src"
+)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE stable-diffusion zip ${CMAKE_THREAD_LIBS_INIT})
 if(SD_WEBP)

examples/cli/README.md

@@ -103,8 +103,9 @@ Generation Options:
   --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
                                            (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
                                            antialiased), or a model name under --hires-upscalers-dir (default: Latent)
-  --extra-sample-args <string>             extra sampler args, key=value list. Currently lcm supports noise_clip_std,
-                                           noise_scale_start, noise_scale_end
+  --extra-sample-args <string>             extra sampler/scheduler args, key=value list. lcm supports noise_clip_std,
+                                           noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift,
+                                           stretch, terminal
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
@@ -160,6 +161,7 @@ Generation Options:
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
   --vae-tiling                             process vae in tiles to reduce memory usage
+  --temporal-tiling                        enable temporal tiling for LTX video VAE decode
   --hires                                  enable highres fix
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
   --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
@@ -169,8 +171,8 @@ Generation Options:
                                            dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
                                            res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] default: euler for Flux/SD3/Wan, euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
-                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
-                                           discrete
+                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2], default:
+                                           model-specific
   --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
                                            "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])

examples/cli/main.cpp

@@ -385,11 +385,32 @@ std::string format_frame_idx(std::string pattern, int frame_idx) {
     return result;
 }
 
+static fs::path get_video_audio_sidecar_path(const SDCliParams& cli_params) {
+    fs::path out_path     = cli_params.output_path;
+    fs::path base_path    = out_path;
+    fs::path ext          = out_path.has_extension() ? out_path.extension() : fs::path{};
+    std::string ext_lower = ext.string();
+    std::transform(ext_lower.begin(), ext_lower.end(), ext_lower.begin(), ::tolower);
+    const EncodedImageFormat output_format = encoded_image_format_from_path(out_path.string());
+    if (!ext.empty()) {
+        if (output_format == EncodedImageFormat::JPEG ||
+            output_format == EncodedImageFormat::PNG ||
+            output_format == EncodedImageFormat::WEBP ||
+            ext_lower == ".avi" ||
+            ext_lower == ".webm") {
+            base_path.replace_extension();
+        }
+    }
+    base_path += ".wav";
+    return base_path;
+}
+
 bool save_results(const SDCliParams& cli_params,
                   const SDContextParams& ctx_params,
                   const SDGenerationParams& gen_params,
                   sd_image_t* results,
-                  int num_results) {
+                  int num_results,
+                  const sd_audio_t* generated_audio = nullptr) {
     if (results == nullptr || num_results <= 0) {
         return false;
     }
@@ -442,6 +463,21 @@ bool save_results(const SDCliParams& cli_params,
         return ok;
     };
 
+    auto write_audio_sidecar = [&](const fs::path& wav_path) {
+        if (generated_audio == nullptr) {
+            return;
+        }
+        if (write_wav_to_file(wav_path.string(),
+                              generated_audio->data,
+                              generated_audio->sample_count,
+                              generated_audio->channels,
+                              generated_audio->sample_rate)) {
+            LOG_INFO("save result audio to '%s'", wav_path.string().c_str());
+        } else {
+            LOG_WARN("failed to save result audio to '%s'", wav_path.string().c_str());
+        }
+    };
+
     int sucessful_reults = 0;
 
     if (std::regex_search(cli_params.output_path, format_specifier_regex)) {
@@ -465,8 +501,16 @@ bool save_results(const SDCliParams& cli_params,
             ext = ".avi";
         fs::path video_path = base_path;
         video_path += ext;
-        if (create_video_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps) == 0) {
+        std::string final_ext_lower = ext.string();
+        std::transform(final_ext_lower.begin(), final_ext_lower.end(), final_ext_lower.begin(), ::tolower);
+        const bool mux_audio = generated_audio != nullptr && (final_ext_lower == ".avi" || final_ext_lower == ".webm");
+        if (create_video_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps, 90, mux_audio ? generated_audio : nullptr) == 0) {
             LOG_INFO("save result video to '%s'", video_path.string().c_str());
+            if (generated_audio != nullptr && !mux_audio) {
+                fs::path wav_path = video_path;
+                wav_path.replace_extension(".wav");
+                write_audio_sidecar(wav_path);
+            }
             return true;
         } else {
             LOG_ERROR("Failed to save result video to '%s'", video_path.string().c_str());
@@ -488,6 +532,9 @@ bool save_results(const SDCliParams& cli_params,
         }
     }
     LOG_INFO("%d/%d images saved", sucessful_reults, num_results);
+    if (generated_audio != nullptr) {
+        write_audio_sidecar(get_video_audio_sidecar_path(cli_params));
+    }
     return sucessful_reults != 0;
 }
 
@@ -701,7 +748,8 @@ 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);
 
     SDImageVec results;
-    int num_results = 0;
+    int num_results             = 0;
+    sd_audio_t* generated_audio = nullptr;
 
     if (cli_params.mode == UPSCALE) {
         num_results = 1;
@@ -733,7 +781,10 @@ int main(int argc, const char* argv[]) {
             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.to_sd_vid_gen_params_t();
-            sd_image_t* generated_video        = generate_video(sd_ctx.get(), &vid_gen_params, &num_results);
+            sd_image_t* generated_video        = nullptr;
+            if (!generate_video(sd_ctx.get(), &vid_gen_params, &generated_video, &num_results, &generated_audio)) {
+                generated_video = nullptr;
+            }
             results.adopt(generated_video, num_results);
         }
 
@@ -775,9 +826,12 @@ int main(int argc, const char* argv[]) {
         }
     }
 
-    if (!save_results(cli_params, ctx_params, gen_params, results.data(), num_results)) {
+    if (!save_results(cli_params, ctx_params, gen_params, results.data(), num_results, generated_audio)) {
+        free_sd_audio(generated_audio);
         return 1;
     }
 
+    free_sd_audio(generated_audio);
+
     return 0;
 }

examples/common/common.cpp

@@ -340,10 +340,18 @@ ArgOptions SDContextParams::get_options() {
          "--high-noise-diffusion-model",
          "path to the standalone high noise diffusion model",
          &high_noise_diffusion_model_path},
+        {"",
+         "--embeddings-connectors",
+         "path to LTXAV embeddings connectors",
+         &embeddings_connectors_path},
         {"",
          "--vae",
          "path to standalone vae model",
          &vae_path},
+        {"",
+         "--audio-vae",
+         "path to standalone LTX audio vae model",
+         &audio_vae_path},
         {"",
          "--taesd",
          "path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)",
@@ -669,7 +677,9 @@ std::string SDContextParams::to_string() const {
         << "  llm_vision_path: \"" << llm_vision_path << "\",\n"
         << "  diffusion_model_path: \"" << diffusion_model_path << "\",\n"
         << "  high_noise_diffusion_model_path: \"" << high_noise_diffusion_model_path << "\",\n"
+        << "  embeddings_connectors_path: \"" << embeddings_connectors_path << "\",\n"
         << "  vae_path: \"" << vae_path << "\",\n"
+        << "  audio_vae_path: \"" << audio_vae_path << "\",\n"
         << "  taesd_path: \"" << taesd_path << "\",\n"
         << "  esrgan_path: \"" << esrgan_path << "\",\n"
         << "  control_net_path: \"" << control_net_path << "\",\n"
@@ -728,7 +738,9 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
         llm_vision_path.c_str(),
         diffusion_model_path.c_str(),
         high_noise_diffusion_model_path.c_str(),
+        embeddings_connectors_path.c_str(),
         vae_path.c_str(),
+        audio_vae_path.c_str(),
         taesd_path.c_str(),
         control_net_path.c_str(),
         embedding_vec.data(),
@@ -821,7 +833,7 @@ ArgOptions SDGenerationParams::get_options() {
          &hires_upscaler},
         {"",
          "--extra-sample-args",
-         "extra sampler args, key=value list. Currently lcm supports noise_clip_std, noise_scale_start, noise_scale_end",
+         "extra sampler/scheduler args, key=value list. lcm supports noise_clip_std, noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift, stretch, terminal",
          &extra_sample_args},
     };
 
@@ -1006,6 +1018,11 @@ ArgOptions SDGenerationParams::get_options() {
          "process vae in tiles to reduce memory usage",
          true,
          &vae_tiling_params.enabled},
+        {"",
+         "--temporal-tiling",
+         "enable temporal tiling for LTX video VAE decode",
+         true,
+         &vae_tiling_params.temporal_tiling},
         {"",
          "--hires",
          "enable highres fix",
@@ -1270,7 +1287,7 @@ ArgOptions SDGenerationParams::get_options() {
          on_high_noise_sample_method_arg},
         {"",
          "--scheduler",
-         "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default: discrete",
+         "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2], default: model-specific",
          on_scheduler_arg},
         {"",
          "--sigmas",
@@ -1703,6 +1720,9 @@ bool SDGenerationParams::from_json_str(
         if (tiling_json.contains("enabled") && tiling_json["enabled"].is_boolean()) {
             vae_tiling_params.enabled = tiling_json["enabled"];
         }
+        if (tiling_json.contains("temporal_tiling") && tiling_json["temporal_tiling"].is_boolean()) {
+            vae_tiling_params.temporal_tiling = tiling_json["temporal_tiling"];
+        }
         if (tiling_json.contains("tile_size_x") && tiling_json["tile_size_x"].is_number_integer()) {
             vae_tiling_params.tile_size_x = tiling_json["tile_size_x"];
         }
@@ -2212,6 +2232,7 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     params.strength                 = strength;
     params.seed                     = seed;
     params.video_frames             = video_frames;
+    params.fps                      = fps;
     params.vace_strength            = vace_strength;
     params.vae_tiling_params        = vae_tiling_params;
     params.cache                    = cache_params;
@@ -2300,6 +2321,7 @@ std::string SDGenerationParams::to_string() const {
         << ", upscale_tile_size: " << hires_upscale_tile_size << " },\n"
         << "  vae_tiling_params: { "
         << vae_tiling_params.enabled << ", "
+        << vae_tiling_params.temporal_tiling << ", "
         << vae_tiling_params.tile_size_x << ", "
         << vae_tiling_params.tile_size_y << ", "
         << vae_tiling_params.target_overlap << ", "

examples/common/common.h

@@ -92,7 +92,9 @@ struct SDContextParams {
     std::string llm_vision_path;
     std::string diffusion_model_path;
     std::string high_noise_diffusion_model_path;
+    std::string embeddings_connectors_path;
     std::string vae_path;
+    std::string audio_vae_path;
     std::string taesd_path;
     std::string esrgan_path;
     std::string control_net_path;
@@ -187,7 +189,7 @@ struct SDGenerationParams {
     int video_frames                     = 1;
     int fps                              = 16;
     float vace_strength                  = 1.f;
-    sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
+    sd_tiling_params_t vae_tiling_params = {false, false, 0, 0, 0.5f, 0.0f, 0.0f};
 
     std::string pm_id_images_dir;
     std::string pm_id_embed_path;

examples/common/media_io.cpp

@@ -613,6 +613,13 @@ typedef struct {
     uint32_t size;
 } avi_index_entry;
 
+typedef struct {
+    char fourcc[4];
+    uint32_t flags;
+    uint32_t offset;
+    uint32_t size;
+} avi_chunk_index_entry;
+
 void write_u32_le(FILE* f, uint32_t val) {
     fwrite(&val, 4, 1, f);
 }
@@ -647,6 +654,33 @@ void write_fourcc(std::vector<uint8_t>& data, const char* fourcc) {
     data.insert(data.end(), fourcc, fourcc + 4);
 }
 
+static std::vector<uint8_t> audio_to_pcm16_bytes(const sd_audio_t* audio) {
+    if (audio == nullptr || audio->data == nullptr || audio->sample_count == 0 || audio->channels == 0 || audio->sample_rate == 0) {
+        return {};
+    }
+
+    const size_t pcm_samples = static_cast<size_t>(audio->sample_count) * static_cast<size_t>(audio->channels);
+    std::vector<uint8_t> bytes(pcm_samples * sizeof(int16_t));
+    auto* pcm = reinterpret_cast<int16_t*>(bytes.data());
+    for (size_t i = 0; i < pcm_samples; ++i) {
+        const float sample = std::clamp(audio->data[i], -1.0f, 1.0f);
+        pcm[i]             = static_cast<int16_t>(std::lrint(sample * 32767.0f));
+    }
+    return bytes;
+}
+
+static std::pair<uint64_t, uint64_t> audio_sample_range_for_video_frame(const sd_audio_t* audio, int frame_idx, int num_frames, int fps) {
+    if (audio == nullptr || fps <= 0 || num_frames <= 0) {
+        return {0, 0};
+    }
+    const uint64_t total = audio->sample_count;
+    const uint64_t start = static_cast<uint64_t>((static_cast<long double>(frame_idx) * total) / num_frames);
+    const uint64_t end   = frame_idx + 1 == num_frames
+                               ? total
+                               : static_cast<uint64_t>((static_cast<long double>(frame_idx + 1) * total) / num_frames);
+    return {start, std::max(start, end)};
+}
+
 EncodedImageFormat encoded_image_format_from_path(const std::string& path) {
     std::string ext = fs::path(path).extension().string();
     std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
@@ -776,7 +810,7 @@ uint8_t* load_image_from_memory(const char* image_bytes,
     return load_image_common(true, image_bytes, len, width, height, expected_width, expected_height, expected_channel);
 }
 
-std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality) {
+std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
     if (num_images == 0) {
         fprintf(stderr, "Error: Image array is empty.\n");
         return {};
@@ -793,7 +827,13 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
     // stb_image_write changes JPEG sampling behavior above quality 90.
     // MJPG AVI playback is more compatible when we keep the encoder on the
     // <= 90 path.
-    const int mjpg_quality = std::clamp(quality, 1, 90);
+    const int mjpg_quality               = std::clamp(quality, 1, 90);
+    const bool has_audio                 = audio != nullptr && audio->data != nullptr && audio->sample_count > 0 && audio->channels > 0 && audio->sample_rate > 0;
+    const std::vector<uint8_t> audio_pcm = audio_to_pcm16_bytes(audio);
+    const uint16_t audio_bits_per_sample = 16;
+    const uint16_t audio_block_align     = has_audio ? static_cast<uint16_t>(audio->channels * (audio_bits_per_sample / 8)) : 0;
+    const uint32_t audio_byte_rate       = has_audio ? static_cast<uint32_t>(audio->sample_rate * audio_block_align) : 0;
+    const uint32_t audio_data_size       = has_audio ? static_cast<uint32_t>(audio_pcm.size()) : 0;
 
     std::vector<uint8_t> avi_data;
     avi_data.reserve(static_cast<size_t>(num_images) * 1024);
@@ -804,7 +844,11 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
     write_fourcc(avi_data, "AVI ");
 
     write_fourcc(avi_data, "LIST");
-    write_u32_le(avi_data, 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40);
+    uint32_t hdrl_size = 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40;
+    if (has_audio) {
+        hdrl_size += 8 + (4 + 8 + 56 + 8 + 16);
+    }
+    write_u32_le(avi_data, hdrl_size);
     write_fourcc(avi_data, "hdrl");
 
     write_fourcc(avi_data, "avih");
@@ -815,7 +859,7 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
     write_u32_le(avi_data, 0x110);
     write_u32_le(avi_data, num_images);
     write_u32_le(avi_data, 0);
-    write_u32_le(avi_data, 1);
+    write_u32_le(avi_data, has_audio ? 2 : 1);
     write_u32_le(avi_data, width * height * 3);
     write_u32_le(avi_data, width);
     write_u32_le(avi_data, height);
@@ -862,12 +906,48 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
     write_u32_le(avi_data, 0);
     write_u32_le(avi_data, 0);
 
+    if (has_audio) {
+        write_fourcc(avi_data, "LIST");
+        write_u32_le(avi_data, 4 + 8 + 56 + 8 + 16);
+        write_fourcc(avi_data, "strl");
+
+        write_fourcc(avi_data, "strh");
+        write_u32_le(avi_data, 56);
+        write_fourcc(avi_data, "auds");
+        write_u32_le(avi_data, 0);
+        write_u32_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+        write_u32_le(avi_data, 0);
+        write_u32_le(avi_data, audio_block_align);
+        write_u32_le(avi_data, audio_byte_rate);
+        write_u32_le(avi_data, 0);
+        write_u32_le(avi_data, static_cast<uint32_t>(audio->sample_count));
+        write_u32_le(avi_data, audio_data_size);
+        write_u32_le(avi_data, static_cast<uint32_t>(-1));
+        write_u32_le(avi_data, audio_block_align);
+        write_u16_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+        write_u16_le(avi_data, 0);
+
+        write_fourcc(avi_data, "strf");
+        write_u32_le(avi_data, 16);
+        write_u16_le(avi_data, 1);
+        write_u16_le(avi_data, static_cast<uint16_t>(audio->channels));
+        write_u32_le(avi_data, audio->sample_rate);
+        write_u32_le(avi_data, audio_byte_rate);
+        write_u16_le(avi_data, audio_block_align);
+        write_u16_le(avi_data, audio_bits_per_sample);
+    }
+
     write_fourcc(avi_data, "LIST");
     const size_t movi_size_pos = avi_data.size();
     write_u32_le(avi_data, 0);
     write_fourcc(avi_data, "movi");
 
-    std::vector<avi_index_entry> index(static_cast<size_t>(num_images));
+    std::vector<avi_chunk_index_entry> index;
+    index.reserve(static_cast<size_t>(num_images) + (has_audio ? 1 : 0));
     std::vector<uint8_t> jpeg_data;
 
     for (int i = 0; i < num_images; i++) {
@@ -884,27 +964,46 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
             return {};
         }
 
-        index[i].offset = static_cast<uint32_t>(avi_data.size());
+        avi_chunk_index_entry video_entry = {};
+        memcpy(video_entry.fourcc, "00dc", 4);
+        video_entry.flags  = 0x10;
+        video_entry.offset = static_cast<uint32_t>(avi_data.size());
         write_fourcc(avi_data, "00dc");
         write_u32_le(avi_data, static_cast<uint32_t>(jpeg_data.size()));
-        index[i].size = (uint32_t)jpeg_data.size();
+        video_entry.size = static_cast<uint32_t>(jpeg_data.size());
         avi_data.insert(avi_data.end(), jpeg_data.begin(), jpeg_data.end());
+        index.push_back(video_entry);
 
         if (jpeg_data.size() % 2) {
             avi_data.push_back(0);
         }
     }
 
+    if (has_audio && !audio_pcm.empty()) {
+        avi_chunk_index_entry audio_entry = {};
+        memcpy(audio_entry.fourcc, "01wb", 4);
+        audio_entry.flags  = 0;
+        audio_entry.offset = static_cast<uint32_t>(avi_data.size());
+        audio_entry.size   = static_cast<uint32_t>(audio_pcm.size());
+        write_fourcc(avi_data, "01wb");
+        write_u32_le(avi_data, static_cast<uint32_t>(audio_pcm.size()));
+        avi_data.insert(avi_data.end(), audio_pcm.begin(), audio_pcm.end());
+        index.push_back(audio_entry);
+        if (audio_pcm.size() % 2 != 0) {
+            avi_data.push_back(0);
+        }
+    }
+
     const size_t movi_size = avi_data.size() - movi_size_pos - 4;
     patch_u32_le(avi_data, movi_size_pos, static_cast<uint32_t>(movi_size));
 
     write_fourcc(avi_data, "idx1");
-    write_u32_le(avi_data, num_images * 16);
-    for (int i = 0; i < num_images; i++) {
-        write_fourcc(avi_data, "00dc");
-        write_u32_le(avi_data, 0x10);
-        write_u32_le(avi_data, index[i].offset);
-        write_u32_le(avi_data, index[i].size);
+    write_u32_le(avi_data, static_cast<uint32_t>(index.size() * 16));
+    for (const auto& entry : index) {
+        write_fourcc(avi_data, entry.fourcc);
+        write_u32_le(avi_data, entry.flags);
+        write_u32_le(avi_data, entry.offset);
+        write_u32_le(avi_data, entry.size);
     }
 
     const size_t file_size = avi_data.size() - riff_size_pos - 4;
@@ -913,8 +1012,8 @@ std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images
     return avi_data;
 }
 
-int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
-    std::vector<uint8_t> avi_data = create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality);
+int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
+    std::vector<uint8_t> avi_data = create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality, audio);
     if (avi_data.empty()) {
         return -1;
     }
@@ -1044,7 +1143,7 @@ int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images
 #endif
 
 #ifdef SD_USE_WEBM
-std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality) {
+std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
     if (num_images == 0) {
         fprintf(stderr, "Error: Image array is empty.\n");
         return {};
@@ -1089,6 +1188,25 @@ std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, in
             video_track->set_display_height(static_cast<uint64_t>(height));
             video_track->set_frame_rate(static_cast<double>(fps));
         }
+
+        uint64_t audio_track_number    = 0;
+        std::vector<uint8_t> audio_pcm = audio_to_pcm16_bytes(audio);
+        if (audio != nullptr && !audio_pcm.empty()) {
+            audio_track_number = segment.AddAudioTrack(static_cast<int32_t>(audio->sample_rate), static_cast<int32_t>(audio->channels), 0);
+            if (audio_track_number == 0) {
+                fprintf(stderr, "Error: Failed to add audio track.\n");
+                return -1;
+            }
+            auto* audio_track = static_cast<mkvmuxer::AudioTrack*>(segment.GetTrackByNumber(audio_track_number));
+            if (audio_track == nullptr) {
+                fprintf(stderr, "Error: Failed to get audio track.\n");
+                return -1;
+            }
+            audio_track->set_codec_id("A_PCM/INT/LIT");
+            audio_track->set_bit_depth(16);
+            audio_track->set_sample_rate(static_cast<double>(audio->sample_rate));
+            audio_track->set_channels(audio->channels);
+        }
         segment.GetSegmentInfo()->set_writing_app("stable-diffusion.cpp");
         segment.GetSegmentInfo()->set_muxing_app("stable-diffusion.cpp");
 
@@ -1118,6 +1236,23 @@ std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, in
                 return -1;
             }
 
+            if (audio_track_number != 0) {
+                auto [audio_begin, audio_end] = audio_sample_range_for_video_frame(audio, i, num_images, fps);
+                const uint64_t frame_samples  = audio_end - audio_begin;
+                if (frame_samples > 0) {
+                    const uint64_t frame_bytes = frame_samples * audio->channels * sizeof(int16_t);
+                    const uint8_t* frame_ptr   = audio_pcm.data() + audio_begin * audio->channels * sizeof(int16_t);
+                    if (!segment.AddFrame(frame_ptr,
+                                          frame_bytes,
+                                          audio_track_number,
+                                          timestamp_ns,
+                                          true)) {
+                        fprintf(stderr, "Error: Failed to mux audio chunk %d into WebM.\n", i);
+                        return -1;
+                    }
+                }
+            }
+
             timestamp_ns += frame_duration_ns;
         }
 
@@ -1133,8 +1268,8 @@ std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, in
     return writer.data();
 }
 
-int create_webm_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
-    std::vector<uint8_t> webm_data = create_webm_from_sd_images_to_vector(images, num_images, fps, quality);
+int create_webm_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
+    std::vector<uint8_t> webm_data = create_webm_from_sd_images_to_vector(images, num_images, fps, quality, audio);
     if (webm_data.empty()) {
         return -1;
     }
@@ -1150,7 +1285,8 @@ std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& ou
                                                            sd_image_t* images,
                                                            int num_images,
                                                            int fps,
-                                                           int quality) {
+                                                           int quality,
+                                                           const sd_audio_t* audio) {
     std::string format = output_format;
     std::transform(format.begin(), format.end(), format.begin(),
                    [](unsigned char c) { return static_cast<char>(tolower(c)); });
@@ -1160,7 +1296,7 @@ std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& ou
 
 #ifdef SD_USE_WEBM
     if (format == "webm") {
-        return create_webm_from_sd_images_to_vector(images, num_images, fps, quality);
+        return create_webm_from_sd_images_to_vector(images, num_images, fps, quality, audio);
     }
 #endif
 
@@ -1170,14 +1306,14 @@ std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& ou
     }
 #endif
 
-    return create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality);
+    return create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality, audio);
 }
 
-int create_video_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+int create_video_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality, const sd_audio_t* audio) {
     std::string path                = filename ? filename : "";
     auto pos                        = path.find_last_of('.');
     std::string ext                 = pos == std::string::npos ? "" : path.substr(pos);
-    std::vector<uint8_t> video_data = create_video_from_sd_images_to_vector(ext, images, num_images, fps, quality);
+    std::vector<uint8_t> video_data = create_video_from_sd_images_to_vector(ext, images, num_images, fps, quality, audio);
     if (video_data.empty()) {
         return -1;
     }
@@ -1187,3 +1323,54 @@ int create_video_from_sd_images(const char* filename, sd_image_t* images, int nu
     }
     return 0;
 }
+
+bool write_wav_to_file(const std::string& path,
+                       const float* interleaved_samples,
+                       uint64_t sample_count,
+                       uint32_t channels,
+                       uint32_t sample_rate) {
+    if (interleaved_samples == nullptr || sample_count == 0 || channels == 0 || sample_rate == 0) {
+        return false;
+    }
+
+    std::ofstream file(path, std::ios::binary);
+    if (!file.is_open()) {
+        return false;
+    }
+
+    uint32_t bits_per_sample  = 16;
+    uint32_t bytes_per_sample = bits_per_sample / 8;
+    uint32_t block_align      = channels * bytes_per_sample;
+    uint32_t byte_rate        = sample_rate * block_align;
+    uint32_t data_size        = static_cast<uint32_t>(sample_count * channels * bytes_per_sample);
+    uint32_t riff_size        = 36 + data_size;
+
+    file.write("RIFF", 4);
+    file.write(reinterpret_cast<const char*>(&riff_size), sizeof(riff_size));
+    file.write("WAVE", 4);
+    file.write("fmt ", 4);
+
+    uint32_t fmt_size            = 16;
+    uint16_t audio_format        = 1;
+    uint16_t wav_channels        = static_cast<uint16_t>(channels);
+    uint16_t wav_block_align     = static_cast<uint16_t>(block_align);
+    uint16_t wav_bits_per_sample = static_cast<uint16_t>(bits_per_sample);
+    file.write(reinterpret_cast<const char*>(&fmt_size), sizeof(fmt_size));
+    file.write(reinterpret_cast<const char*>(&audio_format), sizeof(audio_format));
+    file.write(reinterpret_cast<const char*>(&wav_channels), sizeof(wav_channels));
+    file.write(reinterpret_cast<const char*>(&sample_rate), sizeof(sample_rate));
+    file.write(reinterpret_cast<const char*>(&byte_rate), sizeof(byte_rate));
+    file.write(reinterpret_cast<const char*>(&wav_block_align), sizeof(wav_block_align));
+    file.write(reinterpret_cast<const char*>(&wav_bits_per_sample), sizeof(wav_bits_per_sample));
+
+    file.write("data", 4);
+    file.write(reinterpret_cast<const char*>(&data_size), sizeof(data_size));
+
+    std::vector<int16_t> pcm(sample_count * channels);
+    for (size_t i = 0; i < pcm.size(); ++i) {
+        float sample = std::max(-1.0f, std::min(1.0f, interleaved_samples[i]));
+        pcm[i]       = static_cast<int16_t>(std::lrint(sample * 32767.0f));
+    }
+    file.write(reinterpret_cast<const char*>(pcm.data()), static_cast<std::streamsize>(pcm.size() * sizeof(int16_t)));
+    return file.good();
+}

examples/common/media_io.h

@@ -57,11 +57,13 @@ int create_mjpg_avi_from_sd_images(const char* filename,
                                    sd_image_t* images,
                                    int num_images,
                                    int fps,
-                                   int quality = 90);
+                                   int quality             = 90,
+                                   const sd_audio_t* audio = nullptr);
 std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images,
                                                               int num_images,
                                                               int fps,
-                                                              int quality = 90);
+                                                              int quality             = 90,
+                                                              const sd_audio_t* audio = nullptr);
 
 #ifdef SD_USE_WEBP
 int create_animated_webp_from_sd_images(const char* filename,
@@ -80,22 +82,32 @@ int create_webm_from_sd_images(const char* filename,
                                sd_image_t* images,
                                int num_images,
                                int fps,
-                               int quality = 90);
+                               int quality             = 90,
+                               const sd_audio_t* audio = nullptr);
 std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images,
                                                           int num_images,
                                                           int fps,
-                                                          int quality = 90);
+                                                          int quality             = 90,
+                                                          const sd_audio_t* audio = nullptr);
 #endif
 
 int create_video_from_sd_images(const char* filename,
                                 sd_image_t* images,
                                 int num_images,
                                 int fps,
-                                int quality = 90);
+                                int quality             = 90,
+                                const sd_audio_t* audio = nullptr);
 std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& output_format,
                                                            sd_image_t* images,
                                                            int num_images,
                                                            int fps,
-                                                           int quality = 90);
+                                                           int quality             = 90,
+                                                           const sd_audio_t* audio = nullptr);
+
+bool write_wav_to_file(const std::string& path,
+                       const float* interleaved_samples,
+                       uint64_t sample_count,
+                       uint32_t channels,
+                       uint32_t sample_rate);
 
 #endif  // __MEDIA_IO_H__

examples/server/README.md

@@ -205,8 +205,9 @@ Default Generation Options:
   --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
                                            (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
                                            antialiased), or a model name under --hires-upscalers-dir (default: Latent)
-  --extra-sample-args <string>             extra sampler args, key=value list. Currently lcm supports noise_clip_std,
-                                           noise_scale_start, noise_scale_end
+  --extra-sample-args <string>             extra sampler/scheduler args, key=value list. lcm supports noise_clip_std,
+                                           noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift,
+                                           stretch, terminal
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
@@ -271,8 +272,8 @@ Default Generation Options:
                                            dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
                                            res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] default: euler for Flux/SD3/Wan, euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
-                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
-                                           discrete
+                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2], default:
+                                           model-specific
   --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
                                            "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])

examples/server/async_jobs.cpp

@@ -231,16 +231,21 @@ bool execute_vid_gen_job(ServerRuntime& runtime,
     sd_vid_gen_params_t params = job.vid_gen.to_sd_vid_gen_params_t();
 
     SDImageVec results;
-    int num_results = 0;
+    int num_results             = 0;
+    sd_audio_t* generated_audio = nullptr;
 
     {
         std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
-        sd_image_t* raw_results = generate_video(runtime.sd_ctx, &params, &num_results);
+        sd_image_t* raw_results = nullptr;
+        if (!generate_video(runtime.sd_ctx, &params, &raw_results, &num_results, &generated_audio)) {
+            raw_results = nullptr;
+        }
         results.adopt(raw_results, num_results);
     }
 
     num_results = results.count();
     if (num_results <= 0) {
+        free_sd_audio(generated_audio);
         error_message = "generate_video returned no results";
         return false;
     }
@@ -249,7 +254,9 @@ bool execute_vid_gen_job(ServerRuntime& runtime,
                                                                              results.data(),
                                                                              num_results,
                                                                              job.vid_gen.gen_params.fps,
-                                                                             job.vid_gen.output_compression);
+                                                                             job.vid_gen.output_compression,
+                                                                             generated_audio);
+    free_sd_audio(generated_audio);
     if (video_bytes.empty()) {
         error_message = "failed to encode generated video container";
         return false;