Proposition : Fix TripoSR + Intégration Wan2.1/Applio/SadTalker

Date : 18/01/2026 14:30 Auteur : Claude Status : En attente de validation

Contexte

4 outils IA sur win11 nécessitent des corrections :

1. TripoSR - Cassé (torchmcubes manquant)
2. Wan2.1 - Mode passthrough (pas d'executor)
3. Applio - Mode passthrough (API complexe)
4. SadTalker - Mode passthrough (pas d'executor)

Diagnostic

Plan d'implémentation

Phase 1 : Fix TripoSR (5 min)

# Sur win11
ssh gouro@192.168.1.30
cd I:\TripoSR
venv\Scripts\pip.exe install git+https://github.com/tatsy/torchmcubes.git
# Test
venv\Scripts\python.exe -c "import torchmcubes; print('OK')"

Phase 2 : Ajouter executors dans ai-orchestrator

Fichier : /stock_8to/33800-stack/projects/ai-orchestrator/app/main.py

2.1 Executor Wan2.1

async def execute_wan21_job(params: Dict, job_id: str) -> Dict:
    """Execute Wan2.1 text-to-video job"""
    prompt = params.get("prompt", "")
    resolution = params.get("resolution", "480*832")  # WxH
    steps = params.get("steps", 50)
    guide_scale = params.get("guide_scale", 6.0)
    shift_scale = params.get("shift_scale", 8.0)
    seed = params.get("seed", -1)
    negative_prompt = params.get("negative_prompt", "")

    import aiohttp
    async with aiohttp.ClientSession() as session:
        # Submit via Gradio API
        async with session.post(
            f"http://{WIN11_HOST}:7860/gradio_api/call/t2v_generation",
            json={"data": [prompt, resolution, steps, guide_scale, shift_scale, seed, negative_prompt]},
            timeout=aiohttp.ClientTimeout(total=60)
        ) as resp:
            event_data = await resp.json()
            event_id = event_data.get("event_id")

        # Wait for result (video gen takes time)
        async with session.get(
            f"http://{WIN11_HOST}:7860/gradio_api/call/t2v_generation/{event_id}",
            timeout=aiohttp.ClientTimeout(total=600)  # 10 min max
        ) as resp:
            async for line in resp.content:
                line = line.decode().strip()
                if line.startswith("data: "):
                    data = json.loads(line[6:])
                    if isinstance(data, list) and len(data) > 0:
                        video_info = data[0]
                        # Download and save video
                        # ... (similar to bark executor)

Paramètres simplifiés :

• prompt (str, requis) - Description de la vidéo
• resolution (str, default "480*832") - Format vidéo
• steps (int, default 50) - Qualité
• seed (int, default -1) - Reproductibilité

2.2 Executor SadTalker

async def execute_sadtalker_job(params: Dict, job_id: str) -> Dict:
    """Execute SadTalker talking face job"""
    source_image = params.get("source_image")  # Path or base64
    driven_audio = params.get("driven_audio")  # Path or base64
    preprocess = params.get("preprocess", "crop")  # crop|resize|full
    still_mode = params.get("still_mode", False)
    enhancer = params.get("enhancer", False)  # GFPGAN
    size = params.get("size", 256)  # 256 or 512
    pose_style = params.get("pose_style", 0)  # 0-46

    # Upload files if base64, then call Gradio API
    # ...

Paramètres simplifiés :

• source_image (str, requis) - Image visage
• driven_audio (str, requis) - Audio à synchroniser
• preprocess (str, default "crop")
• enhancer (bool, default False) - GFPGAN upscale

2.3 Executor Applio

async def execute_applio_job(params: Dict, job_id: str) -> Dict:
    """Execute Applio RVC voice conversion job"""
    input_audio = params.get("input_audio")  # Path or base64
    model_name = params.get("model_name")  # Nom du modèle RVC
    pitch = params.get("pitch", 0)  # -12 to +12
    f0_method = params.get("f0_method", "rmvpe")  # rmvpe|crepe|harvest
    index_rate = params.get("index_rate", 0.75)  # 0-1
    protect = params.get("protect", 0.33)  # 0-0.5

    # Call Applio inference endpoint
    # ...

Paramètres simplifiés :

• input_audio (str, requis) - Audio source
• model_name (str, requis) - Modèle voix cible
• pitch (int, default 0) - Ajustement pitch
• f0_method (str, default "rmvpe") - Méthode extraction

2.4 Executor TripoSR

async def execute_triposr_job(params: Dict, job_id: str) -> Dict:
    """Execute TripoSR image-to-3D job"""
    input_image = params.get("input_image")  # Path or base64
    remove_background = params.get("remove_background", True)
    foreground_ratio = params.get("foreground_ratio", 0.85)
    mc_resolution = params.get("mc_resolution", 256)  # 32-320
    output_format = params.get("output_format", "glb")  # obj|glb

    # Call TripoSR Gradio API
    # ...

Paramètres simplifiés :

• input_image (str, requis) - Image objet
• remove_background (bool, default True)
• mc_resolution (int, default 256) - Qualité mesh
• output_format (str, default "glb")

Résumé des modifications

Tests prévus

1. TripoSR : Démarrer via orchestrator, soumettre job image → mesh
2. Wan2.1 : Soumettre job prompt → vidéo
3. SadTalker : Soumettre job image+audio → vidéo talking face
4. Applio : Soumettre job audio+model → audio converti

Risques

• Wan2.1 : Temps de génération long (5-10 min), timeout possible
• Applio : Nécessite modèles RVC pré-téléchargés
• SadTalker : Premier démarrage télécharge checkpoints (~1 Go)

Validation requise

• [ ] Approuver Phase 1 (fix TripoSR)
• [ ] Approuver Phase 2 (executors)
• [ ] Lancer l'implémentation

Temps estimé : 30-45 min total