vaes:
  wan2.1-vae-decoder:
    name: Wan2.1 VAE decoder
    # The complete checkpoint has 126_892_531 parameters. Wan T2V constructs no encoder;
    # retain decoder.*, post_quant_conv.*, and quant_conv.* when counting its loaded subset.
    parameter_count: 73_296_659
    loaded_subset: Decoder, post_quant_conv, and quant_conv; the encoder is not loaded for text-to-video.
    geometry:
      temporal_compression: 4
      spatial_compression: 8
      latent_channels: 16 # Channels in the compressed video.
      output_channels: 3 # Channels in the final RGB frames.
      vae_decoder_channels: 96 # Base channel width inside the decoder.
    # Rough, uncalibrated Wan feature/cache allowance from PR #1978; not a count of RGB frames.
    # feature_cache_bytes ≈ multiplier * batch * vae_decoder_channels * height * width * activation_bytes.
    # Decoded output buffers and resident weights are separate costs; cache/tiling mode affects applicability.
    decode_feature_multiplier: 48
    sources:
      - https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers/tree/0fad780a534b6463e45facd96134c9f345acfa5b/vae
      - https://github.com/hao-ai-lab/FastVideo/blob/81b73f6b88b18803a24510a336b0bae1b0c2c4ef/fastvideo/models/wan/pipeline_config.py
      - https://github.com/hao-ai-lab/FastVideo/blob/81b73f6b88b18803a24510a336b0bae1b0c2c4ef/fastvideo/models/wan/vae.py