Skip to content

VAE encode-decode round-trip is broken in OSS build #13

Description

@soran-ghaderi

On the public action-cond GR-1 checkpoint (2B_GR1_post-train/iter_000050000/model_ema_bf16.pt), model.decode(model.encode(real_video)) produces near-all-black output. PSNR ≈ 7.9 dB vs the input, mean ≈ −0.92 in [−1, 1]. the round-trip is structurally non-identity and is not a quality issue.

The wan2pt1 tokenizer 'cosmos_predict2/_src/predict2/tokenizers/wan2pt1.py lines 594-698' defaults load_mean_std=False, which falls through to video_mean=zeros(...), video_std=ones(...). a no-op outer normalization. The proper statistics live at s3://bucket/cosmos_diffusion_v2/pretrain_weights/tokenizer/wan2pt1/mean_std.pt, which is gated. Without those stats, the encoder emits latents at the wrong scale for the (also un-calibrated) decoder.

The reason it works for the DiT is i think they were trained on this same setup (PSNR ≈ 21 vs real).

Please either:

  • release mean_std.pt (or images_mean_std.pt / video_mean_std.pt) on a public mirror so load_mean_std=True works
  • verify explicitly that the public build supports rollout-decode only and that direct VAE round-trips will not be supported without the gated stats

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions