{
  "format": "split",
  "components": [
    "transformer",
    "text_encoder",
    "vae"
  ],
  "source": "alibaba-pai/CogVideoX-Fun-V1.5-5b-InP",
  "notes": {
    "inpainting": "Transformer patch_embed.proj.weight has 48 input channels (16 latent + 16 masked latent + 16 mask) for inpainting support.",
    "text_encoder": "T5-v1.1-XXL encoder (24 layers, d_model=4096)."
  },
  "recipe": "cogvideox-fun-v1.5-5b-inp",
  "license": "apache-2.0",
  "links": [
    "VideoX-Fun-mlx (inference code): https://github.com/dgrauet/VideoX-Fun-mlx",
    "mlx-forge (conversion tool): https://github.com/dgrauet/mlx-forge",
    "mlx-arsenal (MLX utilities): https://github.com/dgrauet/mlx-arsenal"
  ],
  "usage_url": "https://github.com/dgrauet/VideoX-Fun-mlx",
  "extra_links": [
    "q8 variant: https://huggingface.co/dgrauet/CogVideoX-Fun-V1.5-5b-InP-mlx-q8",
    "q4 variant: https://huggingface.co/dgrauet/CogVideoX-Fun-V1.5-5b-InP-mlx-q4"
  ],
  "quantization_scope": "transformer block Linear weights only, leaving the embeddings, norms and the output projection in bf16",
  "cli_snippet": "pip install mlx sentencepiece pillow numpy huggingface_hub\npip install git+https://github.com/dgrauet/mlx-arsenal.git\ngit clone https://github.com/dgrauet/VideoX-Fun-mlx.git && cd VideoX-Fun-mlx\n\nhuggingface-cli download {repo_id} --local-dir models/cogvideox-fun\n\npython scripts/quick_infer.py \\\n    --model-path models/cogvideox-fun \\\n    --prompt \"a beautiful sunset over the ocean\" \\\n    --output sunset.gif"
}