voice_audio_multi/model_specs/latentsync.json

{
  "schema_version": 1,
  "family": "latentsync",
  "display_name": "LatentSync 1.6",
  "description": "Audio-conditioned lip synchronization of a reference video.",
  "category": "audio_generation",
  "status": "experimental",
  "tasks": [
    "sfx"
  ],
  "modes": [
    "offline"
  ],
  "languages": [
    "auto"
  ],
  "capabilities": {},
  "runtime": {
    "tags": [
      "gguf"
    ]
  },
  "options": {
    "request": [
      {
        "name": "reference_video_file",
        "type": "path",
        "required": true,
        "description": "Reference video containing the face to synchronize. Requires FFmpeg on PATH."
      },
      {
        "name": "num_inference_steps",
        "required": false,
        "type": "int",
        "min": 1,
        "max": 999,
        "default": 20,
        "description": "Number of DDIM denoising steps."
      },
      {
        "name": "cache_interval",
        "required": false,
        "type": "int",
        "min": 1,
        "default": 1,
        "description": "DeepCache refresh interval. One disables caching; three matches the upstream DeepCache preset. Higher values trade output fidelity for speed."
      },
      {
        "name": "guidance_scale",
        "required": false,
        "type": "float",
        "min": 0,
        "default": 1.5,
        "description": "Audio classifier-free guidance scale; values at or below one disable CFG."
      },
      {
        "name": "frames_per_chunk",
        "required": false,
        "type": "int",
        "min": 1,
        "max": 24,
        "default": 16,
        "description": "Frames processed together by the temporal denoiser."
      },
      {
        "name": "fps",
        "required": false,
        "type": "float",
        "min": 1,
        "default": 25,
        "description": "Video sampling and output frame rate."
      },
      {
        "name": "seed",
        "required": false,
        "type": "int",
        "min": 0,
        "default": 1247,
        "description": "Random seed for latent noise and VAE posterior sampling."
      }
    ],
    "session": [
      {
        "name": "weight_type",
        "required": false,
        "type": "enum",
        "values": [
          "native",
          "f32",
          "f16",
          "bf16",
          "q4_0",
          "q4_1",
          "q5_0",
          "q5_1",
          "q2_k",
          "q3_k",
          "q4_k",
          "q5_k",
          "q6_k",
          "q8_0",
          "nvfp4"
        ],
        "default": "native",
        "description": "Weight storage type; native preserves packaged tensor types."
      },
      {
        "name": "vae_cudnn_f16",
        "required": false,
        "type": "bool",
        "default": false,
        "description": "Use cuDNN FP16 VAE execution on CUDA. Requires a build with GGML_CUDA_CUDNN_CONV2D=ON; VAE weights are converted to FP16."
      },
      {
        "name": "unet_cudnn_f16",
        "required": false,
        "type": "bool",
        "default": false,
        "description": "Use cuDNN FP16 convolutions and residual activations in the U-Net on CUDA. Requires a build with GGML_CUDA_CUDNN_CONV2D=ON and FP16 U-Net weights."
      },
      {
        "name": "mem_saver",
        "required": false,
        "type": "bool",
        "default": false,
        "description": "Release components between pipeline stages and run CFG branches sequentially to reduce peak VRAM. Reloading the VAE for decoding adds some latency."
      }
    ],
    "load": []
  },
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "audio-cpp/LatentSync-1.6-GGUF",
      "revision": "main",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "latentsync_1_6_q8_0_f16",
      "display_name": "LatentSync 1.6 Q8/F16",
      "format": "gguf",
      "precision": "q8_0",
      "default": true,
      "target_directory": "LatentSync-1.6-GGUF",
      "files": [
        "latentsync-1.6-q8_0-f16.gguf"
      ]
    },
    {
      "id": "latentsync_1_6_f16",
      "display_name": "LatentSync 1.6 F16",
      "format": "gguf",
      "precision": "f16",
      "default": false,
      "target_directory": "LatentSync-1.6-GGUF",
      "files": [
        "latentsync-1.6-f16.gguf"
      ]
    },
    {
      "id": "latentsync_1_6_f32_f16",
      "display_name": "LatentSync 1.6 F32/F16",
      "format": "gguf",
      "precision": "orig",
      "default": false,
      "target_directory": "LatentSync-1.6-GGUF",
      "files": [
        "latentsync-1.6-f32-f16.gguf"
      ]
    }
  ],
  "dependencies": [],
  "ui": {
    "recommended_package": "latentsync_1_6_q8_0_f16",
    "tags": [
      "GGUF"
    ],
    "docs": [
      "docs/models/latentsync.md"
    ]
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "detector_config": "model:detector_config.json",
        "landmark_config": "model:landmark_config.json"
      },
      "tensors": {
        "weights": {
          "source": "weights:",
          "prefix": "weights"
        }
      }
    }
  ]
}