@qvac/sdk 0.15.0 → 0.16.0

package/dist/schemas/load-model.js

@@ -490,12 +490,14 @@ export const loadClassificationModelRequestSchema = commonModelConfigSchema
 })
     .strict();
 // Custom plugin catch-all: accepts any modelType string EXCEPT built-ins
-export const loadCustomPluginModelRequestSchema = commonModelConfigSchema.extend({
+export const loadCustomPluginModelRequestSchema = commonModelConfigSchema
+    .extend({
     modelType: z.string().refine((val) => !builtInModelTypes.has(val), {
         message: 'Built-in model types must use their specific schema'
     }),
     modelConfig: z.record(z.string(), z.unknown()).optional()
-});
+})
+    .meta({ title: 'LoadModelCustomPluginRequest' });
 // Union of all load model request types (using z.union since each modelType accepts multiple values)
 export const loadModelSrcRequestSchema = z
     .union([
@@ -515,7 +517,8 @@ export const loadModelSrcRequestSchema = z
     .transform((data) => ({
     ...data,
     seed: data.seed ?? false
-}));
+}))
+    .meta({ title: 'LoadModelSrcRequest' });
 // Combined request schema: load new model OR reload config
 export const loadModelRequestSchema = z.union([
     loadModelSrcRequestSchema,

package/dist/schemas/plugin.d.ts

@@ -286,10 +286,10 @@ export declare const PLUGIN_OCR: "@qvac/sdk/ggml-ocr/plugin";
  */
 export declare const PLUGIN_DIFFUSION: "@qvac/sdk/sdcpp-generation/plugin";
 /**
- * Vision-Language-Action plugin (ggml). Supports SmolVLA and π₀.₅ (pi05),
- * dispatched on the GGUF `general.architecture` key.
+ * Vision-Language-Action plugin (ggml). Supports SmolVLA, π₀.₅ (pi05), and
+ * GR00T, dispatched on the GGUF `general.architecture` key.
  * Provides: robot-action inference from one or more camera frames (2 for
- * SmolVLA, 3 for π₀.₅) and a natural-language instruction.
+ * SmolVLA and GR00T, 3 for π₀.₅) and a natural-language instruction.
  */
 export declare const PLUGIN_VLA: "@qvac/sdk/ggml-vla/plugin";
 /**
@@ -326,7 +326,7 @@ export declare const ADDON_TTS: "@qvac/tts-ggml";
 export declare const ADDON_OCR: "@qvac/ocr-ggml";
 /** Native addon package for image generation (stable-diffusion.cpp) */
 export declare const ADDON_DIFFUSION: "@qvac/diffusion-cpp";
-/** Native addon package for vision-language-action inference (SmolVLA / π₀.₅ on ggml) */
+/** Native addon package for vision-language-action inference (SmolVLA / π₀.₅ / GR00T on ggml) */
 export declare const ADDON_VLA: "@qvac/vla-ggml";
 /** Native addon package for image classification (GGML / MobileNetV3) */
 export declare const ADDON_CLASSIFICATION: "@qvac/classification-ggml";

package/dist/schemas/plugin.js

@@ -172,10 +172,10 @@ export const PLUGIN_OCR = '@qvac/sdk/ggml-ocr/plugin';
  */
 export const PLUGIN_DIFFUSION = '@qvac/sdk/sdcpp-generation/plugin';
 /**
- * Vision-Language-Action plugin (ggml). Supports SmolVLA and π₀.₅ (pi05),
- * dispatched on the GGUF `general.architecture` key.
+ * Vision-Language-Action plugin (ggml). Supports SmolVLA, π₀.₅ (pi05), and
+ * GR00T, dispatched on the GGUF `general.architecture` key.
  * Provides: robot-action inference from one or more camera frames (2 for
- * SmolVLA, 3 for π₀.₅) and a natural-language instruction.
+ * SmolVLA and GR00T, 3 for π₀.₅) and a natural-language instruction.
  */
 export const PLUGIN_VLA = '@qvac/sdk/ggml-vla/plugin';
 /**
@@ -226,7 +226,7 @@ export const ADDON_TTS = '@qvac/tts-ggml';
 export const ADDON_OCR = '@qvac/ocr-ggml';
 /** Native addon package for image generation (stable-diffusion.cpp) */
 export const ADDON_DIFFUSION = '@qvac/diffusion-cpp';
-/** Native addon package for vision-language-action inference (SmolVLA / π₀.₅ on ggml) */
+/** Native addon package for vision-language-action inference (SmolVLA / π₀.₅ / GR00T on ggml) */
 export const ADDON_VLA = '@qvac/vla-ggml';
 /** Native addon package for image classification (GGML / MobileNetV3) */
 export const ADDON_CLASSIFICATION = '@qvac/classification-ggml';

package/dist/schemas/reload-config.js

@@ -37,4 +37,6 @@ export const reloadConfigWhisperRequestSchema = reloadConfigRequestBaseSchema.ex
     modelConfig: whisperConfigSchema.partial()
 });
 // Using z.union since modelType accepts multiple values
-export const reloadConfigRequestSchema = z.union([reloadConfigWhisperRequestSchema]);
+export const reloadConfigRequestSchema = z
+    .union([reloadConfigWhisperRequestSchema])
+    .meta({ title: 'ReloadConfigRequest' });

package/dist/schemas/sdcpp-config.d.ts

@@ -10,6 +10,10 @@ export declare const sdcppConfigSchema: z.ZodObject<{
         gpu: "gpu";
         cpu: "cpu";
     }>>;
+    'main-gpu': z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
+        integrated: "integrated";
+        dedicated: "dedicated";
+    }>]>>;
     prediction: z.ZodOptional<z.ZodEnum<{
         auto: "auto";
         eps: "eps";
@@ -278,6 +282,43 @@ export declare const sdcppConfigSchema: z.ZodObject<{
             classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
+    uncondModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
+        src: z.ZodString;
+        name: z.ZodOptional<z.ZodString>;
+        modelId: z.ZodOptional<z.ZodString>;
+        registryPath: z.ZodOptional<z.ZodString>;
+        registrySource: z.ZodOptional<z.ZodString>;
+        blobCoreKey: z.ZodOptional<z.ZodString>;
+        blobIndex: z.ZodOptional<z.ZodNumber>;
+        engine: z.ZodOptional<z.ZodString>;
+        expectedSize: z.ZodOptional<z.ZodNumber>;
+        sha256Checksum: z.ZodOptional<z.ZodString>;
+        addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            embeddings: "embeddings";
+            diffusion: "diffusion";
+            tts: "tts";
+            ocr: "ocr";
+            vla: "vla";
+            "llamacpp-completion": "llamacpp-completion";
+            "whispercpp-transcription": "whispercpp-transcription";
+            "bci-whispercpp-transcription": "bci-whispercpp-transcription";
+            "llamacpp-embedding": "llamacpp-embedding";
+            "nmtcpp-translation": "nmtcpp-translation";
+            "onnx-tts": "onnx-tts";
+            "tts-ggml": "tts-ggml";
+            "parakeet-transcription": "parakeet-transcription";
+            "ggml-ocr": "ggml-ocr";
+            "sdcpp-generation": "sdcpp-generation";
+            "ggml-vla": "ggml-vla";
+            "ggml-classification": "ggml-classification";
+            llm: "llm";
+            whisper: "whisper";
+            bci: "bci";
+            nmt: "nmt";
+            parakeet: "parakeet";
+            classification: "classification";
+        }>, z.ZodLiteral<"vad">]>>;
+    }, z.core.$strip>]>>;
     clipVisionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
         src: z.ZodString;
         name: z.ZodOptional<z.ZodString>;
@@ -315,6 +356,80 @@ export declare const sdcppConfigSchema: z.ZodObject<{
             classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
+    audioVaeModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
+        src: z.ZodString;
+        name: z.ZodOptional<z.ZodString>;
+        modelId: z.ZodOptional<z.ZodString>;
+        registryPath: z.ZodOptional<z.ZodString>;
+        registrySource: z.ZodOptional<z.ZodString>;
+        blobCoreKey: z.ZodOptional<z.ZodString>;
+        blobIndex: z.ZodOptional<z.ZodNumber>;
+        engine: z.ZodOptional<z.ZodString>;
+        expectedSize: z.ZodOptional<z.ZodNumber>;
+        sha256Checksum: z.ZodOptional<z.ZodString>;
+        addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            embeddings: "embeddings";
+            diffusion: "diffusion";
+            tts: "tts";
+            ocr: "ocr";
+            vla: "vla";
+            "llamacpp-completion": "llamacpp-completion";
+            "whispercpp-transcription": "whispercpp-transcription";
+            "bci-whispercpp-transcription": "bci-whispercpp-transcription";
+            "llamacpp-embedding": "llamacpp-embedding";
+            "nmtcpp-translation": "nmtcpp-translation";
+            "onnx-tts": "onnx-tts";
+            "tts-ggml": "tts-ggml";
+            "parakeet-transcription": "parakeet-transcription";
+            "ggml-ocr": "ggml-ocr";
+            "sdcpp-generation": "sdcpp-generation";
+            "ggml-vla": "ggml-vla";
+            "ggml-classification": "ggml-classification";
+            llm: "llm";
+            whisper: "whisper";
+            bci: "bci";
+            nmt: "nmt";
+            parakeet: "parakeet";
+            classification: "classification";
+        }>, z.ZodLiteral<"vad">]>>;
+    }, z.core.$strip>]>>;
+    embeddingsConnectorsModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
+        src: z.ZodString;
+        name: z.ZodOptional<z.ZodString>;
+        modelId: z.ZodOptional<z.ZodString>;
+        registryPath: z.ZodOptional<z.ZodString>;
+        registrySource: z.ZodOptional<z.ZodString>;
+        blobCoreKey: z.ZodOptional<z.ZodString>;
+        blobIndex: z.ZodOptional<z.ZodNumber>;
+        engine: z.ZodOptional<z.ZodString>;
+        expectedSize: z.ZodOptional<z.ZodNumber>;
+        sha256Checksum: z.ZodOptional<z.ZodString>;
+        addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            embeddings: "embeddings";
+            diffusion: "diffusion";
+            tts: "tts";
+            ocr: "ocr";
+            vla: "vla";
+            "llamacpp-completion": "llamacpp-completion";
+            "whispercpp-transcription": "whispercpp-transcription";
+            "bci-whispercpp-transcription": "bci-whispercpp-transcription";
+            "llamacpp-embedding": "llamacpp-embedding";
+            "nmtcpp-translation": "nmtcpp-translation";
+            "onnx-tts": "onnx-tts";
+            "tts-ggml": "tts-ggml";
+            "parakeet-transcription": "parakeet-transcription";
+            "ggml-ocr": "ggml-ocr";
+            "sdcpp-generation": "sdcpp-generation";
+            "ggml-vla": "ggml-vla";
+            "ggml-classification": "ggml-classification";
+            llm: "llm";
+            whisper: "whisper";
+            bci: "bci";
+            nmt: "nmt";
+            parakeet: "parakeet";
+            classification: "classification";
+        }>, z.ZodLiteral<"vad">]>>;
+    }, z.core.$strip>]>>;
     upscaler: z.ZodOptional<z.ZodObject<{
         type: z.ZodOptional<z.ZodLiteral<"esrgan">>;
         model_src: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -364,6 +479,11 @@ export type SdcppConfig = z.input<typeof sdcppConfigSchema>;
 export declare const diffusionStatsSchema: z.ZodObject<{
     modelLoadMs: z.ZodOptional<z.ZodNumber>;
     generationMs: z.ZodOptional<z.ZodNumber>;
+    conditionerMs: z.ZodOptional<z.ZodNumber>;
+    denoiseMs: z.ZodOptional<z.ZodNumber>;
+    vaeMs: z.ZodOptional<z.ZodNumber>;
+    postProcessMs: z.ZodOptional<z.ZodNumber>;
+    stepsPerSecond: z.ZodOptional<z.ZodNumber>;
     totalGenerationMs: z.ZodOptional<z.ZodNumber>;
     totalWallMs: z.ZodOptional<z.ZodNumber>;
     totalSteps: z.ZodOptional<z.ZodNumber>;
@@ -381,6 +501,11 @@ export declare const videoStatsSchema: z.ZodObject<{
     height: z.ZodOptional<z.ZodNumber>;
     modelLoadMs: z.ZodOptional<z.ZodNumber>;
     generationMs: z.ZodOptional<z.ZodNumber>;
+    conditionerMs: z.ZodOptional<z.ZodNumber>;
+    denoiseMs: z.ZodOptional<z.ZodNumber>;
+    vaeMs: z.ZodOptional<z.ZodNumber>;
+    postProcessMs: z.ZodOptional<z.ZodNumber>;
+    stepsPerSecond: z.ZodOptional<z.ZodNumber>;
     totalGenerationMs: z.ZodOptional<z.ZodNumber>;
     totalWallMs: z.ZodOptional<z.ZodNumber>;
     totalSteps: z.ZodOptional<z.ZodNumber>;
@@ -391,6 +516,8 @@ export declare const videoStatsSchema: z.ZodObject<{
     totalVideoFrames: z.ZodOptional<z.ZodNumber>;
     videoFrames: z.ZodOptional<z.ZodNumber>;
     fps: z.ZodOptional<z.ZodNumber>;
+    hasAudio: z.ZodOptional<z.ZodBoolean>;
+    audioSampleRate: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strip>;
 export type VideoStats = z.infer<typeof videoStatsSchema>;
 export declare const diffusionStreamResponseSchema: z.ZodObject<{
@@ -404,6 +531,11 @@ export declare const diffusionStreamResponseSchema: z.ZodObject<{
     stats: z.ZodOptional<z.ZodObject<{
         modelLoadMs: z.ZodOptional<z.ZodNumber>;
         generationMs: z.ZodOptional<z.ZodNumber>;
+        conditionerMs: z.ZodOptional<z.ZodNumber>;
+        denoiseMs: z.ZodOptional<z.ZodNumber>;
+        vaeMs: z.ZodOptional<z.ZodNumber>;
+        postProcessMs: z.ZodOptional<z.ZodNumber>;
+        stepsPerSecond: z.ZodOptional<z.ZodNumber>;
         totalGenerationMs: z.ZodOptional<z.ZodNumber>;
         totalWallMs: z.ZodOptional<z.ZodNumber>;
         totalSteps: z.ZodOptional<z.ZodNumber>;
@@ -430,6 +562,11 @@ export declare const videoStreamResponseSchema: z.ZodObject<{
         height: z.ZodOptional<z.ZodNumber>;
         modelLoadMs: z.ZodOptional<z.ZodNumber>;
         generationMs: z.ZodOptional<z.ZodNumber>;
+        conditionerMs: z.ZodOptional<z.ZodNumber>;
+        denoiseMs: z.ZodOptional<z.ZodNumber>;
+        vaeMs: z.ZodOptional<z.ZodNumber>;
+        postProcessMs: z.ZodOptional<z.ZodNumber>;
+        stepsPerSecond: z.ZodOptional<z.ZodNumber>;
         totalGenerationMs: z.ZodOptional<z.ZodNumber>;
         totalWallMs: z.ZodOptional<z.ZodNumber>;
         totalSteps: z.ZodOptional<z.ZodNumber>;
@@ -440,6 +577,8 @@ export declare const videoStreamResponseSchema: z.ZodObject<{
         totalVideoFrames: z.ZodOptional<z.ZodNumber>;
         videoFrames: z.ZodOptional<z.ZodNumber>;
         fps: z.ZodOptional<z.ZodNumber>;
+        hasAudio: z.ZodOptional<z.ZodBoolean>;
+        audioSampleRate: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strip>>;
 }, z.core.$strip>;
 export type VideoStreamResponse = z.infer<typeof videoStreamResponseSchema>;
@@ -640,6 +779,105 @@ export declare const videoRequestSchema: z.ZodObject<{
     vae_tiling: z.ZodOptional<z.ZodBoolean>;
     vae_tile_size: z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>>;
     vae_tile_overlap: z.ZodOptional<z.ZodNumber>;
+    temporal_tiling: z.ZodOptional<z.ZodBoolean>;
+    cache_mode: z.ZodOptional<z.ZodEnum<{
+        disabled: "disabled";
+        easycache: "easycache";
+        ucache: "ucache";
+        dbcache: "dbcache";
+        taylorseer: "taylorseer";
+        "cache-dit": "cache-dit";
+    }>>;
+    cache_preset: z.ZodOptional<z.ZodString>;
+    cache_threshold: z.ZodOptional<z.ZodNumber>;
+    mode: z.ZodEnum<{
+        txt2vid: "txt2vid";
+        img2vid: "img2vid";
+    }>;
+    init_image: z.ZodOptional<z.ZodString>;
+    strength: z.ZodOptional<z.ZodNumber>;
+}, z.core.$strip>;
+export declare const ltxVideoRequestSchema: z.ZodObject<{
+    modelId: z.ZodString;
+    requestId: z.ZodOptional<z.ZodString>;
+    prompt: z.ZodString;
+    negative_prompt: z.ZodOptional<z.ZodString>;
+    width: z.ZodOptional<z.ZodNumber>;
+    height: z.ZodOptional<z.ZodNumber>;
+    video_frames: z.ZodOptional<z.ZodNumber>;
+    fps: z.ZodOptional<z.ZodNumber>;
+    seed: z.ZodOptional<z.ZodNumber>;
+    steps: z.ZodOptional<z.ZodNumber>;
+    sampling_method: z.ZodOptional<z.ZodEnum<{
+        euler: "euler";
+        euler_a: "euler_a";
+        heun: "heun";
+        dpm2: "dpm2";
+        "dpm++2m": "dpm++2m";
+        "dpm++2mv2": "dpm++2mv2";
+        "dpm++2s_a": "dpm++2s_a";
+        lcm: "lcm";
+        ipndm: "ipndm";
+        ipndm_v: "ipndm_v";
+        ddim_trailing: "ddim_trailing";
+        tcd: "tcd";
+        res_multistep: "res_multistep";
+        res_2s: "res_2s";
+    }>>;
+    scheduler: z.ZodOptional<z.ZodEnum<{
+        lcm: "lcm";
+        discrete: "discrete";
+        karras: "karras";
+        exponential: "exponential";
+        ays: "ays";
+        gits: "gits";
+        sgm_uniform: "sgm_uniform";
+        simple: "simple";
+        smoothstep: "smoothstep";
+        kl_optimal: "kl_optimal";
+        bong_tangent: "bong_tangent";
+    }>>;
+    cfg_scale: z.ZodOptional<z.ZodNumber>;
+    flow_shift: z.ZodOptional<z.ZodNumber>;
+    high_noise_steps: z.ZodOptional<z.ZodNumber>;
+    high_noise_sampler: z.ZodOptional<z.ZodEnum<{
+        euler: "euler";
+        euler_a: "euler_a";
+        heun: "heun";
+        dpm2: "dpm2";
+        "dpm++2m": "dpm++2m";
+        "dpm++2mv2": "dpm++2mv2";
+        "dpm++2s_a": "dpm++2s_a";
+        lcm: "lcm";
+        ipndm: "ipndm";
+        ipndm_v: "ipndm_v";
+        ddim_trailing: "ddim_trailing";
+        tcd: "tcd";
+        res_multistep: "res_multistep";
+        res_2s: "res_2s";
+    }>>;
+    high_noise_scheduler: z.ZodOptional<z.ZodEnum<{
+        lcm: "lcm";
+        discrete: "discrete";
+        karras: "karras";
+        exponential: "exponential";
+        ays: "ays";
+        gits: "gits";
+        sgm_uniform: "sgm_uniform";
+        simple: "simple";
+        smoothstep: "smoothstep";
+        kl_optimal: "kl_optimal";
+        bong_tangent: "bong_tangent";
+    }>>;
+    high_noise_cfg_scale: z.ZodOptional<z.ZodNumber>;
+    high_noise_flow_shift: z.ZodOptional<z.ZodNumber>;
+    moe_boundary: z.ZodOptional<z.ZodNumber>;
+    vace_strength: z.ZodOptional<z.ZodNumber>;
+    control_frames: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    vae_tiling: z.ZodOptional<z.ZodBoolean>;
+    vae_tile_size: z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>>;
+    vae_tile_overlap: z.ZodOptional<z.ZodNumber>;
+    temporal_tiling: z.ZodOptional<z.ZodBoolean>;
     cache_mode: z.ZodOptional<z.ZodEnum<{
         disabled: "disabled";
         easycache: "easycache";
@@ -738,6 +976,7 @@ export declare const videoStreamRequestSchema: z.ZodObject<{
     vae_tiling: z.ZodOptional<z.ZodBoolean>;
     vae_tile_size: z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>>;
     vae_tile_overlap: z.ZodOptional<z.ZodNumber>;
+    temporal_tiling: z.ZodOptional<z.ZodBoolean>;
     cache_mode: z.ZodOptional<z.ZodEnum<{
         disabled: "disabled";
         easycache: "easycache";

package/dist/schemas/sdcpp-config.js

@@ -46,17 +46,35 @@ export const sdcppConfigSchema = z.object({
         .default('diffusion')
         .describe('Operation mode for the diffusion plugin. ' +
         "`'diffusion'` (default) builds a full SD / SDXL / SD3 / FLUX pipeline from " +
-        'the primary model plus optional auxiliary text encoders, VAE, and ESRGAN ' +
+        'the primary model plus optional auxiliary text encoders, VAE, unconditional ' +
+        'diffusion model, and ESRGAN ' +
         'upscaler, and exposes diffusion({ ... }). ' +
         "`'upscale'` builds a standalone ESRGAN upscaler from the primary model " +
         'file alone (auxiliary model sources are ignored) and exposes upscale({ ... }). ' +
-        "`'video'` builds a Wan `VideoStableDiffusion` pipeline and exposes video({ ... }). " +
+        "`'video'` builds a `VideoStableDiffusion` pipeline and exposes video({ ... }). " +
+        'The video layout is selected from the auxiliary sources: supplying ' +
+        '`embeddingsConnectorsModelSrc` loads the LTX-2 layout (Gemma text encoder ' +
+        'via `llmModelSrc` + video VAE + connectors, optional `audioVaeModelSrc` for ' +
+        'synchronized audio); otherwise the Wan layout is used (UMT5 text encoder ' +
+        'via `t5XxlModelSrc` + VAE). ' +
         'On React Native, loading the video model on-device will likely fail ' +
         'because the video diffusion models currently ' +
         'shipped by the SDK are too large to load on typical mobile devices; ' +
         'pass a `delegate` to `loadModel(...)` to run generation on a desktop peer instead.'),
     threads: z.number().optional(),
     device: z.enum(['gpu', 'cpu']).optional(),
+    'main-gpu': z
+        .union([z.number().int().min(0), z.enum(['integrated', 'dedicated'])])
+        .optional()
+        .describe('GPU to pin when `device` is "gpu": a GPU-device index, "integrated", or ' +
+        '"dedicated" (the discrete GPU with the most VRAM). Omit to let the ' +
+        'backend choose the first enumerated device. Resolved against the ' +
+        "addon's own ggml device enumeration, so it cannot desync from the " +
+        'device list the backend actually uses. If an explicit request cannot ' +
+        'be satisfied (e.g. "integrated" with no integrated GPU, "dedicated" ' +
+        'with no discrete GPU, or an out-of-range index) the addon falls back ' +
+        'to CPU rather than substituting a different GPU. Stripped on mobile ' +
+        '(single-GPU devices).'),
     prediction: z
         .enum(['auto', 'eps', 'v', 'edm_v', 'flow', 'flux2_flow'])
         .optional()
@@ -116,17 +134,37 @@ export const sdcppConfigSchema = z.object({
         .describe('T5-XXL text encoder model — required for SD3'),
     llmModelSrc: modelSrcInputSchema
         .optional()
-        .describe('LLM text encoder model (e.g. Qwen3) — required for FLUX.2 [klein]'),
+        .describe('LLM text encoder model — required for FLUX.2 [klein] (Qwen3), ' +
+        'Ideogram 4 (Qwen3-VL), and LTX-2 video (Gemma).'),
     vaeModelSrc: modelSrcInputSchema
         .optional()
-        .describe('VAE decoder model — required for FLUX.2 [klein], optional for SDXL'),
+        .describe('VAE decoder model — required for FLUX.2 [klein], Ideogram 4, and ' +
+        'LTX-2 video (video VAE); optional for SDXL.'),
     highNoiseDiffusionModelSrc: modelSrcInputSchema
         .optional()
         .describe('High-noise diffusion expert — required for Wan 2.2 mixture-of-experts video models'),
+    uncondModelSrc: modelSrcInputSchema
+        .optional()
+        .describe('Unconditional diffusion model — Ideogram 4 only. Requires diffusion mode, ' +
+        'llmModelSrc (Qwen3-VL), vaeModelSrc, and a JSON-serialized structured ' +
+        'caption with explicit bounding boxes as the generation prompt. Plain-text ' +
+        "prompts produce degenerate output or the model's placeholder response."),
     clipVisionModelSrc: modelSrcInputSchema
         .optional()
         .describe('OpenCLIP ViT-H/14 weights (`clip_vision_h.safetensors`). Required for ' +
-        'Wan image-to-video (`img2vid`); omit for text-to-video-only pipelines.'),
+        'Wan image-to-video (`img2vid`); omit for text-to-video-only pipelines. ' +
+        'Not used by LTX-2 (its img2vid path needs no CLIP-vision projection).'),
+    audioVaeModelSrc: modelSrcInputSchema
+        .optional()
+        .describe('Audio VAE decoder model — LTX-2 video only. Enables the synchronized ' +
+        '48 kHz audio track muxed into the output AVI; omit for silent video. ' +
+        'Ignored by the Wan layout.'),
+    embeddingsConnectorsModelSrc: modelSrcInputSchema
+        .optional()
+        .describe('Text-embedding connector weights — required for LTX-2 video. Its ' +
+        'presence selects the LTX-2 video layout (Gemma text encoder via ' +
+        '`llmModelSrc` + video VAE via `vaeModelSrc` + these connectors) instead ' +
+        'of the Wan layout.'),
     upscaler: z
         .object({
         type: z
@@ -182,7 +220,27 @@ export const diffusionStatsSchema = z.object({
     generationMs: z
         .number()
         .optional()
-        .describe('Wall-clock time in milliseconds spent generating images.'),
+        .describe('Wall-clock time in milliseconds spent generating the output.'),
+    conditionerMs: z
+        .number()
+        .optional()
+        .describe('Time in milliseconds spent conditioning the prompt before denoising.'),
+    denoiseMs: z
+        .number()
+        .optional()
+        .describe('Time in milliseconds spent in the diffusion denoising loop.'),
+    vaeMs: z
+        .number()
+        .optional()
+        .describe('Time in milliseconds spent decoding diffusion latents with the VAE.'),
+    postProcessMs: z
+        .number()
+        .optional()
+        .describe('Time in milliseconds spent encoding, upscaling, muxing, and emitting outputs.'),
+    stepsPerSecond: z
+        .number()
+        .optional()
+        .describe('Diffusion denoising throughput in sampling steps per second.'),
     totalGenerationMs: z
         .number()
         .optional()
@@ -209,6 +267,11 @@ export const videoStatsSchema = diffusionStatsSchema
     .pick({
     modelLoadMs: true,
     generationMs: true,
+    conditionerMs: true,
+    denoiseMs: true,
+    vaeMs: true,
+    postProcessMs: true,
+    stepsPerSecond: true,
     totalGenerationMs: true,
     totalWallMs: true,
     totalSteps: true,
@@ -226,7 +289,16 @@ export const videoStatsSchema = diffusionStatsSchema
     fps: z
         .number()
         .optional()
-        .describe('Frames-per-second metadata for the most recent generated video.')
+        .describe('Frames-per-second metadata for the most recent generated video.'),
+    hasAudio: z
+        .boolean()
+        .optional()
+        .describe('True when the output AVI includes a muxed audio track (LTX-2 loaded ' +
+        'with audioVaeModelSrc), false otherwise.'),
+    audioSampleRate: z
+        .number()
+        .optional()
+        .describe('Sample rate (Hz) of the muxed audio track; 0 when there is no audio.')
 });
 export const diffusionStreamResponseSchema = z.object({
     type: z.literal('diffusionStream'),
@@ -395,14 +467,16 @@ const videoGenerationBaseSchema = z.object({
         .positive()
         .multipleOf(16)
         .optional()
-        .describe('Video width in pixels (must be a multiple of 16).'),
+        .describe('Video width in pixels (must be a multiple of 16). LTX-2 additionally ' +
+        'requires a multiple of 32, validated against the loaded model before generation.'),
     height: z
         .number()
         .int()
         .positive()
         .multipleOf(16)
         .optional()
-        .describe('Video height in pixels (must be a multiple of 16).'),
+        .describe('Video height in pixels (must be a multiple of 16). LTX-2 additionally ' +
+        'requires a multiple of 32, validated against the loaded model before generation.'),
     video_frames: z
         .number()
         .int()
@@ -410,7 +484,9 @@ const videoGenerationBaseSchema = z.object({
         message: 'video_frames must be an integer >= 5 of the form (4*k + 1)'
     })
         .optional()
-        .describe('Frame count for the generated video; must satisfy (4*k + 1), where k>=1.'),
+        .describe('Frame count for the generated video; must satisfy (4*k + 1), where k>=1. ' +
+        'LTX-2 additionally requires the stricter (8*k + 1) with a max of 257, ' +
+        'validated against the loaded model before generation.'),
     fps: z
         .number()
         .positive()
@@ -474,6 +550,11 @@ const videoGenerationBaseSchema = z.object({
         .optional()
         .describe('VAE tile size override.'),
     vae_tile_overlap: z.number().optional().describe('VAE tile overlap override.'),
+    temporal_tiling: z
+        .boolean()
+        .optional()
+        .describe('LTX-2 only: tile the video VAE decode along the time axis to cap peak ' +
+        'VRAM for HD / long clips. No effect on Wan (spatial-only VAE).'),
     cache_mode: cacheModeSchema.optional().describe('Step-caching algorithm.'),
     cache_preset: z
         .string()
@@ -527,6 +608,31 @@ function refineVideoMode(data, ctx) {
     }
 }
 export const videoRequestSchema = videoRequestObjectSchema.superRefine(refineVideoMode);
+function refineLtxVideoRequest(data, ctx) {
+    if (data.width !== undefined && data.width % 32 !== 0) {
+        ctx.addIssue({
+            code: 'custom',
+            path: ['width'],
+            message: 'LTX-2 width must be a multiple of 32.'
+        });
+    }
+    if (data.height !== undefined && data.height % 32 !== 0) {
+        ctx.addIssue({
+            code: 'custom',
+            path: ['height'],
+            message: 'LTX-2 height must be a multiple of 32.'
+        });
+    }
+    if (data.video_frames !== undefined &&
+        (data.video_frames > 257 || (data.video_frames - 1) % 8 !== 0)) {
+        ctx.addIssue({
+            code: 'custom',
+            path: ['video_frames'],
+            message: 'LTX-2 video_frames must be at most 257 and satisfy (8*k + 1), where k>=1.'
+        });
+    }
+}
+export const ltxVideoRequestSchema = videoRequestSchema.superRefine(refineLtxVideoRequest);
 export const videoStreamRequestSchema = videoRequestObjectSchema
     .extend({ type: z.literal('videoStream') })
     .superRefine(refineVideoMode);