Skip to content

LTX-2 unified VAE decoder strict loading fails when per-channel statistics are absent #36

Description

@andyelka-creator

Summary

Image-to-video generation with Lightricks/LTX-2 now reaches Stage 1 and then fails while loading the unified VAE decoder for latent upsampling.

The unified Hugging Face VAE safetensors do not include per_channel_statistics.mean or per_channel_statistics.std. The decoder module initializes identity defaults for these values, but strict load still requires them to be present in the sanitized weights.

Reproduction

Run distilled I2V generation with Lightricks/LTX-2 and a source image. After Stage 1 completes, generation loads vae/decoder from the unified VAE checkpoint.

Actual result

ValueError: Missing 2 parameters:
per_channel_statistics.mean,
per_channel_statistics.std.

Expected result

The decoder should keep its identity per-channel statistics defaults when the unified checkpoint omits these optional statistics, allowing strict loading to continue.

Impact

Distilled I2V generation cannot proceed from Stage 1 latent upsampling to Stage 2 decoding with the current Lightricks/LTX-2 unified VAE checkpoint.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions