{
  "schema_version": 1,
  "family": "liveavatar",
  "display_name": "LiveAvatar",
  "description": "LiveAvatar audio-driven avatar video generation using the Wan2.2 S2V architecture.",
  "category": "audio_generation",
  "status": "experimental",
  "tasks": [
    "sfx"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "auto"
  ],
  "runtime": {
    "tags": [
      "cuda",
      "gguf",
      "stream"
    ]
  },
  "capabilities": {
    "sfx": [
      "prompt_generation"
    ]
  },
  "options": {
    "request": [
      {
        "name": "reference_image_file",
        "type": "string",
        "description": "Reference image path used as the identity/appearance condition.",
        "required": true
      },
      {
        "name": "generation_mode",
        "type": "enum",
        "values": [
          "offline",
          "liveavatar"
        ],
        "description": "Use whole-clip generation or LiveAvatar blockwise generation.",
        "required": false,
        "default": "offline"
      },
      {
        "name": "height",
        "type": "int",
        "description": "Generated frame height; must be a positive multiple of 16.",
        "required": false,
        "min": 16,
        "default": 480
      },
      {
        "name": "width",
        "type": "int",
        "description": "Generated frame width; must be a positive multiple of 16.",
        "required": false,
        "min": 16,
        "default": 832
      },
      {
        "name": "num_frames",
        "type": "int",
        "description": "Frames generated per LiveAvatar segment; must be positive and a multiple of 4 in blockwise mode (streaming or generation_mode=liveavatar).",
        "required": false,
        "min": 1,
        "default": 48
      },
      {
        "name": "num_clips",
        "type": "int",
        "description": "Maximum LiveAvatar clips to generate from the input audio.",
        "required": false,
        "min": 1,
        "default": 1
      },
      {
        "name": "num_inference_steps",
        "type": "int",
        "description": "Denoising steps; offline generation uses UniPC, blockwise generation uses Euler. The LiveAvatar DMD adapter is designed for 4 steps.",
        "required": false,
        "min": 1,
        "default": 4
      },
      {
        "name": "guidance_scale",
        "type": "float",
        "description": "Classifier-free guidance scale; LiveAvatar blockwise generation requires 0.",
        "required": false,
        "min": 0.0,
        "default": 0.0
      },
      {
        "name": "batched_cfg",
        "type": "bool",
        "description": "Run classifier-free guidance in one denoiser graph.",
        "required": false,
        "default": true
      },
      {
        "name": "sage_attention",
        "type": "bool",
        "description": "Use CUDA SageAttention when the backend supports it.",
        "required": false,
        "default": true
      },
      {
        "name": "mem_saver",
        "type": "bool",
        "description": "Reuse immutable condition data and temporary denoiser graph storage across sequential blocks.",
        "required": false,
        "default": true
      },
      {
        "name": "denoiser_layerwise",
        "type": "bool",
        "description": "Run the denoiser in layerwise mode. Offline generation requires batched_cfg=true and guidance_scale > 1.",
        "required": false,
        "default": false
      },
      {
        "name": "denoiser_layerwise_batch",
        "type": "int",
        "description": "Layer group size used by layerwise denoising.",
        "required": false,
        "min": 1,
        "default": 16
      },
      {
        "name": "vae_encoder_chunk_size",
        "type": "int",
        "description": "Cached VAE encoder chunk size.",
        "required": false,
        "min": 1,
        "default": 16
      },
      {
        "name": "vae_decoder_tile_size",
        "type": "int",
        "description": "Opt-in VAE decoder spatial tile size for large outputs; 0 disables tiling. Enabled tile sizes must be multiples of 8 and larger than the 64-pixel overlap.",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "target_cache_blocks",
        "type": "int",
        "description": "Target KV cache window in block units; 0 keeps the full clip cache.",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "vae_cache_f16",
        "type": "bool",
        "description": "Store cached VAE encoder state in F16.",
        "required": false,
        "default": true
      },
      {
        "name": "shift",
        "type": "float",
        "description": "Flow scheduler shift; must be > 0.",
        "required": false,
        "min": 0.0,
        "default": 3.0
      },
      {
        "name": "negative_prompt",
        "type": "string",
        "description": "Negative prompt for classifier-free guidance.",
        "required": false
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Generation seed.",
        "required": false,
        "min": 0,
        "default": 420
      }
    ],
    "session": [
      {
        "name": "denoiser_weight_streaming",
        "type": "bool",
        "description": "Keep denoiser transformer blocks in host memory and stage layer groups on the accelerator.",
        "required": false,
        "default": false
      }
    ],
    "load": [
      {
        "name": "support_gguf",
        "type": "string",
        "description": "Support GGUF containing text_encoder/audio_encoder tensors and the embedded tokenizer sidecar.",
        "required": false,
        "default": "Wan2.2-S2V-Support-Q4_K_S-F16.gguf"
      },
      {
        "name": "vae_gguf",
        "type": "string",
        "description": "Wan video VAE component GGUF file relative to the model root.",
        "required": false,
        "default": "Wan2.2-S2V-VAE-F16.gguf"
      },
      {
        "name": "denoiser_gguf",
        "type": "string",
        "description": "Denoiser DiT GGUF file relative to the model root.",
        "required": false,
        "default": "Wan2.2-S2V-14B-NVFP4-LORA.gguf"
      }
    ]
  },
  "dependencies": [],
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "audio-cpp/LiveAvatar-GGUF",
      "revision": "main",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "liveavatar_nvfp4_lora",
      "display_name": "LiveAvatar NVFP4 LoRA GGUF",
      "default": true,
      "format": "gguf",
      "precision": "native",
      "target_directory": "LiveAvatar-GGUF",
      "files": [
        "Wan2.2-S2V-Support-Q4_K_S-F16.gguf",
        "Wan2.2-S2V-VAE-F16.gguf",
        "Wan2.2-S2V-14B-NVFP4-LORA.gguf"
      ]
    }
  ],
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": "."
      },
      "files": {
        "tokenizer_model": "model:tokenizer/spiece.model"
      },
      "tensors": {
        "text_encoder_weights": {
          "source": "model:Wan2.2-S2V-Support-Q4_K_S-F16.gguf",
          "prefix": "text_encoder"
        },
        "audio_encoder_weights": {
          "source": "model:Wan2.2-S2V-Support-Q4_K_S-F16.gguf",
          "prefix": "audio_encoder"
        }
      }
    },
    {
      "format": "gguf",
      "roots": {
        "model": "."
      },
      "tensors": {
        "vae_weights": {
          "source": "model:Wan2.2-S2V-VAE-F16.gguf",
          "prefix": "vae"
        }
      }
    },
    {
      "format": "gguf",
      "roots": {
        "model": "."
      },
      "tensors": {
        "denoiser_weights": {
          "source": "model:Wan2.2-S2V-14B-NVFP4-LORA.gguf"
        }
      }
    }
  ],
  "ui": {
    "recommended_package": "liveavatar_nvfp4_lora",
    "tags": [
      "SFX",
      "GGUF",
      "Stream"
    ],
  "docs": [
    "docs/models/liveavatar.md"
  ]
  }
}