@qvac/sdk 0.15.0 → 0.16.0

package/NOTICE

@@ -127,6 +127,12 @@ Third-Party Model Licenses
     https://huggingface.co/qvac/MedPsy-4B-GGUF
   nb-tiny-ggml-model
     https://huggingface.co/NbAiLab/nb-whisper-tiny
+  parler-indic-f16
+    https://huggingface.co/ai4bharat/indic-parler-tts
+  parler-large-v1-f16
+    https://huggingface.co/parler-tts/parler-tts-large-v1
+  parler-mini-v1-f16
+    https://huggingface.co/parler-tts/parler-tts-mini-v1
   parseq
     https://github.com/felixdittrich92/OnnxTR
   pi05-base-q-aggressive
@@ -147,6 +153,8 @@ Third-Party Model Licenses
     https://huggingface.co/Qwen/Qwen3-4B-GGUF
   Qwen3-8B-GGUF
     https://huggingface.co/Qwen/Qwen3-8B-GGUF
+  Qwen3-Embedding-0.6B-Q8_0
+    https://huggingface.co/Qwen/Qwen3-Embedding-0.6B
   Qwen3-VL-2B-Instruct-GGUF
     https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct-GGUF
   Qwen3.5-0.8B-GGUF
@@ -243,14 +251,6 @@ Third-Party Model Licenses
 
   AfriqueGemma-4B-GGUF
     https://huggingface.co/mradermacher/AfriqueGemma-4B-GGUF
-  decoder_joint-model
-    https://huggingface.co/istupakov/parakeet-tdt-0.6b-v3-onnx
-  diar_streaming_sortformer_4spk-v2-onnx
-    https://huggingface.co/cgus/diar_streaming_sortformer_4spk-v2-onnx
-  encoder-model
-    https://huggingface.co/istupakov/parakeet-tdt-0.6b-v3-onnx
-  parakeet-ctc-0.6b-ONNX
-    https://huggingface.co/onnx-community/parakeet-ctc-0.6b-ONNX
   parakeet-ctc-0.6b.f16
     https://huggingface.co/nvidia/parakeet-ctc-0.6b
   parakeet-ctc-0.6b.q4_0
@@ -263,18 +263,12 @@ Third-Party Model Licenses
     https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
   parakeet-eou-120m-v1.q8_0
     https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
-  parakeet-rs
-    https://huggingface.co/altunenes/parakeet-rs
-  parakeet-tdt-0.6b-v3-onnx
-    https://huggingface.co/istupakov/parakeet-tdt-0.6b-v3-onnx
   parakeet-tdt-0.6b-v3.f16
     https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
   parakeet-tdt-0.6b-v3.q4_0
     https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
   parakeet-tdt-0.6b-v3.q8_0
     https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
-  preprocessor
-    https://huggingface.co/istupakov/parakeet-tdt-0.6b-v3-onnx
   sortformer-4spk-v1.f16
     https://huggingface.co/nvidia/diar_sortformer_4spk-v1
   sortformer-4spk-v1.q4_0
@@ -305,6 +299,14 @@ Third-Party Model Licenses
 
 --- mit (MIT License) ---
 
+  acestep-5Hz-lm-0.6B-Q8_0
+    https://huggingface.co/ACE-Step/acestep-5Hz-lm-0.6B
+  acestep-v15-sft-Q8_0
+    https://huggingface.co/ACE-Step/acestep-v15-sft
+  acestep-v15-turbo-Q4_K_M
+    https://huggingface.co/ACE-Step/Ace-Step1.5
+  acestep-v15-turbo-Q8_0
+    https://huggingface.co/ACE-Step/Ace-Step1.5
   bitnet_b1_58-3B-TQ2_0
     https://huggingface.co/1bitLLM/bitnet_b1_58-3B
   bitnet_b1_58-large-TQ2_0
@@ -313,8 +315,6 @@ Third-Party Model Licenses
     https://huggingface.co/1bitLLM/bitnet_b1_58-xl
   Cangjie5_TC
     https://github.com/Jackchows/Cangjie5
-  chatterbox-multilingual-ONNX
-    https://huggingface.co/onnx-community/chatterbox-multilingual-ONNX
   chatterbox-s3gen
     https://huggingface.co/ResembleAI/chatterbox-turbo
   chatterbox-s3gen-mtl
@@ -323,20 +323,8 @@ Third-Party Model Licenses
     https://huggingface.co/ResembleAI/chatterbox
   chatterbox-t3-turbo
     https://huggingface.co/ResembleAI/chatterbox-turbo
-  chatterbox-turbo-ONNX
-    https://huggingface.co/ResembleAI/chatterbox-turbo-ONNX
   de-base-ggml-model-f16
     https://huggingface.co/aware-ai/whisper-base-german
-  denoiser_core_legacy_fixed63
-    s3:///qvac_models_compiled/onnx/lavasr/2026-04-03/denoiser_core_legacy_fixed63.onnx
-  enhancer_backbone
-    s3:///qvac_models_compiled/onnx/lavasr/2026-04-03/enhancer_backbone.onnx
-  enhancer_backbone.onnx
-    s3:///qvac_models_compiled/onnx/lavasr/2026-04-03/enhancer_backbone.onnx.data
-  enhancer_spec_head
-    s3:///qvac_models_compiled/onnx/lavasr/2026-04-03/enhancer_spec_head.onnx
-  enhancer_spec_head.onnx
-    s3:///qvac_models_compiled/onnx/lavasr/2026-04-03/enhancer_spec_head.onnx.data
   es-tiny-ggml-model-f16
     https://huggingface.co/JaepaX/whisper-tiny-es
   ggml-indictrans2-en-indic-1B-f16
@@ -365,6 +353,8 @@ Third-Party Model Licenses
     https://voiceinput.futo.org/VoiceInput/tiny_en_acft_q8_0.bin
   tiny_en_acft_q8_0
     https://voiceinput.futo.org/VoiceInput/tiny_en_acft_q8_0.bin
+  vae-BF16
+    https://huggingface.co/ACE-Step/Ace-Step1.5
   whisper-vad
     https://huggingface.co/ggml-org/whisper-vad
   whisper.cpp
@@ -585,21 +575,17 @@ Third-Party Model Licenses
     https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2.1
   diar_streaming_sortformer_4spk-v2.1.q8_0
     https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2.1
+  groot-q8_vf16
+    https://huggingface.co/nvidia/GR00T-N1.7-LIBERO
 
 --- openrail ---
 
   supertonic
     s3:///qvac_models_compiled/supertonic/2026-05-08/supertonic.gguf
-  supertonic
-    https://huggingface.co/Supertone/supertonic
-  supertonic-2
-    https://huggingface.co/Supertone/supertonic-2
   supertonic-q4_0
     s3:///qvac_models_compiled/ggml/supertonic/2026-05-18/supertonic-q4_0.gguf
   supertonic-q8_0
     s3:///qvac_models_compiled/ggml/supertonic/2026-05-18/supertonic-q8_0.gguf
-  Supertonic-TTS-ONNX
-    https://huggingface.co/onnx-community/Supertonic-TTS-ONNX
   supertonic2
     s3:///qvac_models_compiled/supertonic/2026-05-08/supertonic2.gguf
   supertonic2-q4_0
@@ -637,17 +623,17 @@ JavaScript Dependencies
 
   @hyperswarm/secret-stream@6.9.1
     https://github.com/holepunchto/hyperswarm-secret-stream
-  @qvac/bci-whispercpp@0.4.1
+  @qvac/bci-whispercpp@0.5.1
     https://github.com/tetherto/qvac
-  @qvac/classification-ggml@0.10.2
+  @qvac/classification-ggml@0.13.0
     https://github.com/tetherto/qvac
   @qvac/decoder-audio@0.5.0
     https://github.com/tetherto/qvac
   @qvac/diagnostics@0.1.2
     https://github.com/tetherto/qvac
-  @qvac/diffusion-cpp@0.13.3
+  @qvac/diffusion-cpp@0.15.0
     https://github.com/tetherto/qvac
-  @qvac/embed-llamacpp@0.26.2
+  @qvac/embed-llamacpp@0.28.0
     https://github.com/tetherto/qvac
   @qvac/error@0.1.1
   @qvac/infer-base@0.4.2
@@ -656,10 +642,10 @@ JavaScript Dependencies
     https://github.com/tetherto/qvac
   @qvac/langdetect-text@0.1.2
     https://github.com/tetherto/qvac
-  @qvac/llm-llamacpp@0.36.3
+  @qvac/llm-llamacpp@0.38.0
     https://github.com/tetherto/qvac
   @qvac/logging@0.1.0
-  @qvac/ocr-ggml@0.10.2
+  @qvac/ocr-ggml@0.12.1
     https://github.com/tetherto/qvac
   @qvac/rag@0.6.4
     https://github.com/tetherto/qvac
@@ -667,15 +653,15 @@ JavaScript Dependencies
     https://github.com/tetherto/qvac
   @qvac/registry-schema@0.3.0
     https://github.com/tetherto/qvac
-  @qvac/transcription-parakeet@0.9.1
+  @qvac/transcription-parakeet@0.10.1
     https://github.com/tetherto/qvac
-  @qvac/transcription-whispercpp@0.11.1
+  @qvac/transcription-whispercpp@0.12.1
     https://github.com/tetherto/qvac
-  @qvac/translation-nmtcpp@7.2.2
+  @qvac/translation-nmtcpp@8.1.0
     https://github.com/tetherto/qvac
-  @qvac/tts-ggml@0.4.2
+  @qvac/tts-ggml@0.5.1
     https://github.com/tetherto/qvac
-  @qvac/vla-ggml@0.11.2
+  @qvac/vla-ggml@0.14.0
     https://github.com/tetherto/qvac
   adaptive-timeout@1.0.1
     https://github.com/holepunchto/adaptive-timeout

package/dist/client/api/ocr.d.ts

@@ -10,6 +10,15 @@ import { type OCRClientParams, type OCRTextBlock, type OCRStats } from '../../sc
  * @returns Object with blockStream generator, blocks promise, and stats promise
  * @example
  * ```typescript
+ * // Load an OCR model first. Registry GGUF models resolve their pipeline
+ * // and detector automatically:
+ * //   EasyOCR (default): OCR_LATIN  (CRAFT detector auto-derived)
+ * //   DocTR:             OCR_DOCTR  (DBNet detector auto-derived)
+ * const modelId = await loadModel({
+ *   modelSrc: OCR_LATIN.src,
+ *   modelType: MODEL_TYPES.ggmlOcr,
+ * });
+ *
  * // Non-streaming mode (default) - get all blocks at once
  * const { blocks } = ocr({ modelId, image: "/path/to/image.png" });
  * for (const block of await blocks) {

package/dist/client/api/ocr.js

@@ -11,6 +11,15 @@ import { stream as streamRpc } from '../rpc/rpc-client.js';
  * @returns Object with blockStream generator, blocks promise, and stats promise
  * @example
  * ```typescript
+ * // Load an OCR model first. Registry GGUF models resolve their pipeline
+ * // and detector automatically:
+ * //   EasyOCR (default): OCR_LATIN  (CRAFT detector auto-derived)
+ * //   DocTR:             OCR_DOCTR  (DBNet detector auto-derived)
+ * const modelId = await loadModel({
+ *   modelSrc: OCR_LATIN.src,
+ *   modelType: MODEL_TYPES.ggmlOcr,
+ * });
+ *
  * // Non-streaming mode (default) - get all blocks at once
  * const { blocks } = ocr({ modelId, image: "/path/to/image.png" });
  * for (const block of await blocks) {

package/dist/client/api/translate.js

@@ -1,7 +1,5 @@
 import { stream as streamRpc } from '../rpc/rpc-client.js';
-import { translateResponseSchema, normalizeModelType, ModelType } from '../../schemas/index.js';
-import { detectOne } from '@qvac/langdetect-text';
-import { TranslationFailedError } from '../../utils/errors-client.js';
+import { translateResponseSchema } from '../../schemas/index.js';
 /**
  * Translates text from one language to another using a specified translation model.
  * Supports both NMT (Neural Machine Translation) and LLM models.
@@ -44,22 +42,13 @@ import { TranslationFailedError } from '../../utils/errors-client.js';
  * ```
  */
 export function translate(params, options) {
-    const canonicalModelType = normalizeModelType(params.modelType);
-    const isLlm = canonicalModelType === ModelType.llamacppCompletion;
-    let sourceLanguage = isLlm ? params.from : undefined;
-    if (!sourceLanguage && isLlm) {
-        const detected = detectOne(params.text);
-        if (detected.code === 'und' || detected.language === 'Undetermined') {
-            throw new TranslationFailedError("Could not detect the source language. Please specify the 'from' parameter explicitly.");
-        }
-        sourceLanguage = detected.language;
-    }
+    // Source-language auto-detection lives in the worker now
+    // (server/bare/ops/translate.ts): `from` is passed through when present and
+    // resolved server-side when absent, so every language binding shares one
+    // detector instead of each shipping its own.
     const request = {
         type: 'translate',
-        ...params,
-        ...(isLlm && {
-            from: sourceLanguage
-        })
+        ...params
     };
     let stats;
     let statsResolver = () => { };

package/dist/client/api/video.d.ts

@@ -16,10 +16,13 @@ export interface VideoResult {
  * @param params - Video request parameters (model, prompt, dimensions, frame count, fps, sampler, seed, etc.).
  * @returns A result object exposing `requestId` (stable identifier for this in-flight generation), `progressStream` (async iterator of `{ step, totalSteps, elapsedMs }`), `outputs` (promise of the generated video buffers, typically a single AVI file), and `stats` (promise of generation statistics).
  *
- * Supports `txt2vid` (text-to-video) and `img2vid` (image-to-video). For `img2vid`,
- * load the Wan pipeline with `modelConfig.clipVisionModelSrc` set to
- * `clip_vision_h.safetensors`. On React Native, prefer a `modelId` loaded with a
- * `delegate` since the bundled video diffusion models are too large for typical mobile devices.
+ * Supports `txt2vid` (text-to-video) and `img2vid` (image-to-video) for both the
+ * Wan and LTX-2 layouts. For Wan `img2vid`, load the pipeline with
+ * `modelConfig.clipVisionModelSrc` set to `clip_vision_h.safetensors`; LTX-2
+ * `img2vid` conditions on the first frame through its video VAE and needs no
+ * CLIP vision weights (the same LTX-2 model loaded for txt2vid also does
+ * img2vid). On React Native, prefer a `modelId` loaded with a `delegate` since
+ * the bundled video diffusion models are too large for typical mobile devices.
  *
  * @example Basic txt2vid generation
  * ```typescript

package/dist/client/api/video.js

@@ -8,10 +8,13 @@ import { decodeBase64, encodeBase64 } from '../../utils/encoding.js';
  * @param params - Video request parameters (model, prompt, dimensions, frame count, fps, sampler, seed, etc.).
  * @returns A result object exposing `requestId` (stable identifier for this in-flight generation), `progressStream` (async iterator of `{ step, totalSteps, elapsedMs }`), `outputs` (promise of the generated video buffers, typically a single AVI file), and `stats` (promise of generation statistics).
  *
- * Supports `txt2vid` (text-to-video) and `img2vid` (image-to-video). For `img2vid`,
- * load the Wan pipeline with `modelConfig.clipVisionModelSrc` set to
- * `clip_vision_h.safetensors`. On React Native, prefer a `modelId` loaded with a
- * `delegate` since the bundled video diffusion models are too large for typical mobile devices.
+ * Supports `txt2vid` (text-to-video) and `img2vid` (image-to-video) for both the
+ * Wan and LTX-2 layouts. For Wan `img2vid`, load the pipeline with
+ * `modelConfig.clipVisionModelSrc` set to `clip_vision_h.safetensors`; LTX-2
+ * `img2vid` conditions on the first frame through its video VAE and needs no
+ * CLIP vision weights (the same LTX-2 model loaded for txt2vid also does
+ * img2vid). On React Native, prefer a `modelId` loaded with a `delegate` since
+ * the bundled video diffusion models are too large for typical mobile devices.
  *
  * @example Basic txt2vid generation
  * ```typescript

package/dist/client/api/vla.d.ts

@@ -1,32 +1,36 @@
 import { type VlaClientRunParams, type VlaClientRunResult, type VlaHparams } from '../../schemas/index';
 /**
- * Run VLA inference on a loaded model (SmolVLA or π₀.₅) and return the
+ * Run VLA inference on a loaded model (SmolVLA, π₀.₅, or GR00T) and return the
  * produced action chunk plus per-stage timings.
  *
  * @param params - Inference inputs.
  * @param params.modelId - Identifier of the loaded VLA model (returned by
  *   `loadModel({ modelType: "vla", ... })`).
- * @param params.images - The preprocessed camera frames; each is a
- *   `Float32Array` of length `3 * imgWidth * imgHeight` in CHW layout, range
- *   `[-1, 1]`. Pass exactly `hparams.numCameras` frames (2 for SmolVLA, 3
- *   for π₀.₅). Use the addon's `preprocessImage()` (re-exported as
- *   `vlaPreprocessImage`) to produce them.
+ * @param params.images - The preprocessed camera frames, one per camera. For
+ *   pixel-input models (SmolVLA, π₀.₅) each is a `Float32Array` of length
+ *   `3 * imgWidth * imgHeight` in CHW layout, range `[-1, 1]`, produced by the
+ *   addon's `preprocessImage()` (re-exported as `vlaPreprocessImage`). For
+ *   patch-input models (GR00T, `hparams.imageInputMode === 'patches'`) each is
+ *   the pre-patchified buffer of length `hparams.imagePatchElems` instead. Pass
+ *   exactly `hparams.numCameras` frames (2 for SmolVLA and GR00T, 3 for π₀.₅).
  * @param params.imgWidth - Width of each preprocessed image; must equal
  *   `hparams.visionImageSize`.
  * @param params.imgHeight - Height of each preprocessed image; must equal
  *   `hparams.visionImageSize`.
  * @param params.state - Robot end-effector / gripper state. For
- *   continuous-state models (SmolVLA) pad to `hparams.maxStateDim` with
+ *   continuous-state models (SmolVLA, GR00T) pad to `hparams.maxStateDim` with
  *   `vlaPadState`. For discrete-state models (π₀.₅,
  *   `hparams.stateInputMode === 'discrete'`) the state is tokenised into the
  *   prompt and this buffer is ignored — pass an empty `Float32Array(0)`.
  * @param params.tokens - Tokenized instruction (`Int32Array` of length
  *   `hparams.tokenizerMaxLength`). Tokenize on the consumer side with the
- *   model's tokenizer (SmolVLM2 for SmolVLA, PaliGemma/Gemma for π₀.₅).
+ *   model's tokenizer (SmolVLM2 for SmolVLA, PaliGemma/Gemma for π₀.₅,
+ *   Qwen3-VL for GR00T).
  * @param params.mask - Token attention mask (`Uint8Array` matching `tokens`).
- * @param params.noise - Optional seeded noise prior
+ * @param params.noise - Seeded noise prior
  *   (`Float32Array` of length `hparams.chunkSize * hparams.maxActionDim`).
- *   When omitted the addon samples its own prior.
+ *   Optional for SmolVLA / π₀.₅ (addon samples its own prior when omitted).
+ *   Required for GR00T (`hparams.imageInputMode === 'patches'`).
  * @returns A `VlaClientRunResult` with the produced `actions` Float32Array
  *   (length `chunkSize * actionDim`), the corresponding `chunkSize` /
  *   `actionDim` returned by the addon, and optional per-stage `stats`.

package/dist/client/api/vla.js

@@ -13,33 +13,37 @@ function f32FromBase64(b64) {
     return new Float32Array(bytes.buffer, bytes.byteOffset, bytes.byteLength / 4);
 }
 /**
- * Run VLA inference on a loaded model (SmolVLA or π₀.₅) and return the
+ * Run VLA inference on a loaded model (SmolVLA, π₀.₅, or GR00T) and return the
  * produced action chunk plus per-stage timings.
  *
  * @param params - Inference inputs.
  * @param params.modelId - Identifier of the loaded VLA model (returned by
  *   `loadModel({ modelType: "vla", ... })`).
- * @param params.images - The preprocessed camera frames; each is a
- *   `Float32Array` of length `3 * imgWidth * imgHeight` in CHW layout, range
- *   `[-1, 1]`. Pass exactly `hparams.numCameras` frames (2 for SmolVLA, 3
- *   for π₀.₅). Use the addon's `preprocessImage()` (re-exported as
- *   `vlaPreprocessImage`) to produce them.
+ * @param params.images - The preprocessed camera frames, one per camera. For
+ *   pixel-input models (SmolVLA, π₀.₅) each is a `Float32Array` of length
+ *   `3 * imgWidth * imgHeight` in CHW layout, range `[-1, 1]`, produced by the
+ *   addon's `preprocessImage()` (re-exported as `vlaPreprocessImage`). For
+ *   patch-input models (GR00T, `hparams.imageInputMode === 'patches'`) each is
+ *   the pre-patchified buffer of length `hparams.imagePatchElems` instead. Pass
+ *   exactly `hparams.numCameras` frames (2 for SmolVLA and GR00T, 3 for π₀.₅).
  * @param params.imgWidth - Width of each preprocessed image; must equal
  *   `hparams.visionImageSize`.
  * @param params.imgHeight - Height of each preprocessed image; must equal
  *   `hparams.visionImageSize`.
  * @param params.state - Robot end-effector / gripper state. For
- *   continuous-state models (SmolVLA) pad to `hparams.maxStateDim` with
+ *   continuous-state models (SmolVLA, GR00T) pad to `hparams.maxStateDim` with
  *   `vlaPadState`. For discrete-state models (π₀.₅,
  *   `hparams.stateInputMode === 'discrete'`) the state is tokenised into the
  *   prompt and this buffer is ignored — pass an empty `Float32Array(0)`.
  * @param params.tokens - Tokenized instruction (`Int32Array` of length
  *   `hparams.tokenizerMaxLength`). Tokenize on the consumer side with the
- *   model's tokenizer (SmolVLM2 for SmolVLA, PaliGemma/Gemma for π₀.₅).
+ *   model's tokenizer (SmolVLM2 for SmolVLA, PaliGemma/Gemma for π₀.₅,
+ *   Qwen3-VL for GR00T).
  * @param params.mask - Token attention mask (`Uint8Array` matching `tokens`).
- * @param params.noise - Optional seeded noise prior
+ * @param params.noise - Seeded noise prior
  *   (`Float32Array` of length `hparams.chunkSize * hparams.maxActionDim`).
- *   When omitted the addon samples its own prior.
+ *   Optional for SmolVLA / π₀.₅ (addon samples its own prior when omitted).
+ *   Required for GR00T (`hparams.imageInputMode === 'patches'`).
  * @returns A `VlaClientRunResult` with the produced `actions` Float32Array
  *   (length `chunkSize * actionDim`), the corresponding `chunkSize` /
  *   `actionDim` returned by the addon, and optional per-stage `stats`.

package/dist/client/rpc/rpc-error.js

@@ -1,4 +1,4 @@
-import { ContextOverflowError, RequestIdConflictError, RequestNotFoundError, RequestRejectedByPolicyError } from '../../utils/errors-server.js';
+import { ContextOverflowError, RequestIdConflictError, RequestNotFoundError, RequestRejectedByPolicyError, TranslationFailedError } from '../../utils/errors-server.js';
 export class RPCError extends Error {
     timestamp;
     remoteStack;
@@ -112,6 +112,12 @@ const RECONSTRUCTORS = {
     },
     CONTEXT_OVERFLOW: (response) => {
         return new ContextOverflowError(readOptionalNumberField(response.typedFields, 'promptTokens'), readOptionalNumberField(response.typedFields, 'ctxSize'), readOptionalStringField(response.typedFields, 'modelId'), response.cause);
+    },
+    // Source-language autodetect now fails server-side (translate detection moved
+    // into the worker), so this crosses the RPC and reconstructs as the typed
+    // class -- matching the Python client's reconstructor.
+    TRANSLATION_FAILED: (response) => {
+        return new TranslationFailedError(response.message, response.cause);
     }
 };
 /**

package/dist/examples/diffusion-ideogram.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=diffusion-ideogram.d.ts.map

package/dist/examples/diffusion-ideogram.js

@@ -0,0 +1,108 @@
+import { diffusion, loadModel, unloadModel } from '@qvac/sdk';
+import fs from 'fs';
+import path from 'path';
+const diffusionModelSrc = process.argv[2] ||
+    'https://huggingface.co/leejet/ideogram-4-GGUF/resolve/main/ideogram4-Q4_0.gguf';
+const uncondModelSrc = process.argv[3] ||
+    'https://huggingface.co/leejet/ideogram-4-GGUF/resolve/main/ideogram4_uncond-Q4_0.gguf';
+const llmModelSrc = process.argv[4] ||
+    'https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/resolve/main/Qwen3-VL-8B-Instruct-Q4_K_M.gguf';
+const vaeModelSrc = process.argv[5] ||
+    'https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/vae/flux2-vae.safetensors';
+const outputDir = process.argv[6] || '.';
+function formatMb(bytes) {
+    return (bytes / 1e6).toFixed(1);
+}
+// Ideogram 4 expects a JSON-serialized structured caption with explicit
+// bounding boxes. Plain-text prompts produce degenerate or placeholder output.
+const structuredPrompt = {
+    high_level_description: 'A bright product photo of one yellow coffee mug on a clean office desk with a tiny readable label that says IDEO.',
+    style_description: {
+        aesthetics: 'simple product photography, clean bright office desk, minimal composition',
+        lighting: 'soft studio light with gentle shadows',
+        photo: 'sharp focus, crisp readable label text',
+        medium: 'product photo',
+        color_palette: ['#FFD54A', '#FFFFFF', '#111111', '#7EC8E3']
+    },
+    compositional_deconstruction: {
+        canvas: 'Square canvas, upright orientation. All text is horizontal and readable.',
+        background: 'Clean white office desk with a soft blue background gradient.',
+        elements: [
+            {
+                type: 'obj',
+                bbox: [230, 260, 800, 740],
+                desc: 'Exactly one matte yellow ceramic coffee mug centered on the desk.'
+            },
+            {
+                type: 'text',
+                bbox: [510, 360, 620, 640],
+                text: 'IDEO',
+                desc: 'Small black label printed horizontally on the mug.'
+            }
+        ]
+    }
+};
+if (process.argv.includes('--help')) {
+    console.log('Usage: tsx examples/diffusion-ideogram.ts [model] [uncond] [llm] [vae] [output-dir]');
+    console.log('▸ Defaults use the four Hugging Face model URLs supported by the diffusion addon.');
+    process.exit(0);
+}
+let modelId;
+try {
+    console.log('▸ Loading Ideogram 4 split-layout model...');
+    modelId = await loadModel({
+        modelSrc: diffusionModelSrc,
+        modelType: 'sdcpp-generation',
+        modelConfig: {
+            device: 'gpu',
+            threads: 4,
+            diffusion_fa: true,
+            offload_to_cpu: true,
+            llmModelSrc,
+            vaeModelSrc,
+            uncondModelSrc
+        },
+        onProgress: (progress) => {
+            const line = `▸ Downloading ${progress.percentage.toFixed(0)}% (${formatMb(progress.downloaded)}/${formatMb(progress.total)} MB)`;
+            process.stderr.write(process.stderr.isTTY ? `\r${line}` : `${line}\n`);
+            if (progress.percentage >= 100)
+                process.stderr.write('\n');
+        }
+    });
+    console.log(`▸ Model loaded: ${modelId}`);
+    const { progressStream, outputs, stats } = diffusion({
+        modelId,
+        prompt: JSON.stringify(structuredPrompt),
+        width: 768,
+        height: 768,
+        steps: 16,
+        cfg_scale: 7,
+        seed: 42
+    });
+    for await (const { step, totalSteps } of progressStream) {
+        console.log(`▸ step ${step}/${totalSteps}`);
+    }
+    const buffers = await outputs;
+    for (let i = 0; i < buffers.length; i++) {
+        const outputPath = path.join(outputDir, `ideogram_${i}.png`);
+        fs.writeFileSync(outputPath, buffers[i]);
+        console.log(`▸ Saved ${outputPath}`);
+    }
+    console.log('▸ Stats:', await stats);
+    await unloadModel({ modelId, clearStorage: false });
+    modelId = undefined;
+    console.log('▸ Done');
+    process.exit(0);
+}
+catch (error) {
+    if (modelId) {
+        try {
+            await unloadModel({ modelId, clearStorage: false });
+        }
+        catch {
+            // Preserve the original error.
+        }
+    }
+    console.error('✖', error);
+    process.exit(1);
+}

package/dist/examples/diffusion-img2vid-ltx.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=diffusion-img2vid-ltx.d.ts.map

package/dist/examples/diffusion-img2vid-ltx.js

@@ -0,0 +1,98 @@
+import { loadModel, unloadModel, video } from '@qvac/sdk';
+import fs from 'fs';
+import path from 'path';
+// Image-to-video (+ synchronized audio) with LTX-2. Unlike Wan I2V, LTX-2
+// conditions on the first frame through its video VAE, so it needs NO CLIP
+// vision weights (clipVisionModelSrc) — the same LTX-2 layout used for
+// txt2vid works for img2vid, you just add an init_image at generation time.
+//
+// This example needs powerful hardware: LTX-2.3 is a ~21B model, budget at
+// least 20+ GB of video / unified memory.
+//
+// No LTX-2 model constants ship in the SDK registry, so the model sources
+// default to public HuggingFace URLs (the SDK downloads http(s) sources
+// directly). Pass your own paths / URLs via argv to override.
+const HF = 'https://huggingface.co';
+const initImagePath = process.argv[2];
+const prompt = process.argv[3] || 'the subject slowly turns and smiles, soft natural lighting, cinematic';
+const outputDir = process.argv[4] || '.';
+const diffusionModelSrc = process.argv[5] ||
+    `${HF}/QuantStack/LTX-2.3-GGUF/resolve/main/LTX-2.3-distilled-1.1/LTX-2.3-22B-distilled-1.1-Q5_K_M.gguf`;
+const llmModelSrc = process.argv[6] || `${HF}/unsloth/gemma-3-12b-it-GGUF/resolve/main/gemma-3-12b-it-UD-Q4_K_XL.gguf`;
+const vaeModelSrc = process.argv[7] ||
+    `${HF}/unsloth/LTX-2.3-GGUF/resolve/main/vae/ltx-2.3-22b-distilled_video_vae.safetensors`;
+const audioVaeModelSrc = process.argv[8] ||
+    `${HF}/unsloth/LTX-2.3-GGUF/resolve/main/vae/ltx-2.3-22b-distilled_audio_vae.safetensors`;
+const embeddingsConnectorsModelSrc = process.argv[9] ||
+    `${HF}/unsloth/LTX-2.3-GGUF/resolve/main/text_encoders/ltx-2.3-22b-distilled_embeddings_connectors.safetensors`;
+if (!initImagePath) {
+    console.error('✖ init image path is required');
+    console.error('Usage: bun run bare:example dist/examples/diffusion-img2vid-ltx.js ' +
+        '<initImagePath> [prompt] [outputDir] ' +
+        '[diffusionModelSrc] [llmModelSrc] [vaeModelSrc] [audioVaeModelSrc] [embeddingsConnectorsModelSrc]');
+    process.exit(1);
+}
+try {
+    console.log('▸ Loading LTX-2 model (diffusion + Gemma + video VAE + audio VAE + connectors)...');
+    // no clipVisionModelSrc — LTX-2 img2vid does not use CLIP vision.
+    const modelId = await loadModel({
+        modelSrc: diffusionModelSrc,
+        modelType: 'sdcpp-generation',
+        modelConfig: {
+            mode: 'video',
+            device: 'gpu',
+            threads: 4,
+            // Supplying embeddingsConnectorsModelSrc selects the LTX-2 layout.
+            llmModelSrc,
+            vaeModelSrc,
+            audioVaeModelSrc,
+            embeddingsConnectorsModelSrc,
+            diffusion_fa: true,
+            vae_tiling: true
+        },
+        onProgress: (p) => {
+            const mb = (n) => (n / 1e6).toFixed(1);
+            const line = `▸ Downloading ${p.percentage.toFixed(0)}% (${mb(p.downloaded)}/${mb(p.total)} MB)`;
+            process.stderr.write(process.stderr.isTTY ? `\r${line}` : `${line}\n`);
+            if (p.percentage >= 100)
+                process.stderr.write('\n');
+        }
+    });
+    console.log(`▸ Model loaded: ${modelId}`);
+    const init_image = new Uint8Array(fs.readFileSync(initImagePath));
+    console.log(`▸ Generating video for: "${prompt}"`);
+    const { progressStream, outputs, stats } = video({
+        modelId,
+        mode: 'img2vid',
+        prompt,
+        init_image,
+        negative_prompt: 'blurry, distorted, low quality, jittery, static, frozen, distorted audio',
+        strength: 0.85,
+        // LTX-2 requires width/height as multiples of 32 and (8*k + 1) frames (max 257).
+        width: 512,
+        height: 320,
+        video_frames: 121,
+        fps: 24,
+        steps: 10,
+        cfg_scale: 1.0,
+        seed: 42,
+        temporal_tiling: true
+    });
+    for await (const { step, totalSteps } of progressStream) {
+        console.log(`▸ step ${step}/${totalSteps}`);
+    }
+    const buffers = await outputs;
+    for (let i = 0; i < buffers.length; i++) {
+        const outputPath = path.join(outputDir, `ltx_i2v_${i}.avi`);
+        fs.writeFileSync(outputPath, buffers[i]);
+        console.log(`▸ Saved ${outputPath} (play in VLC to hear the muxed audio track)`);
+    }
+    console.log('▸ Stats:', await stats);
+    await unloadModel({ modelId, clearStorage: false });
+    console.log('▸ Done');
+    process.exit(0);
+}
+catch (error) {
+    console.error('✖', error);
+    process.exit(1);
+}

package/dist/examples/diffusion-txt2vid-ltx.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=diffusion-txt2vid-ltx.d.ts.map