High failure rate in T2AV generation

[illimit3]

Describe
When using the MOVA pipeline for pure T2V/T2AV generation, I am experiencing a highly unstable generation process. Approximately 40% of the generated videos result in corrupted, solid-color outputs (the entire video is just a flat color with no coherent structures or details).


Following the standard T2V approach for MOVA, I am passing a pure white PIL.Image as the image condition to the pipeline. I strongly suspect the issue lies in how the pipeline_mova.py loads, encodes, or concatenates this pure white frame in the prepare_latents stage, leading to a latent collapse during the diffusion process.

[illimit3]

usually all drak frames in the whole video, a few pieces of characters may occur at the end of the video with less than 1 second