THUDM / CogVideo

text and image to video generation: CogVideoX (2024) and CogVideo (ICLR 2023)
Apache License 2.0
9.46k stars 891 forks source link

Generated video with third-party watermarks #378

Open yclin925 opened 2 months ago

yclin925 commented 2 months ago

System Info / 系統信息

As tile, The generated videos appears as a "pond5" watermark. Is this normal?

Information / 问题信息

Reproduction / 复现过程

  1. Deployment of CogVideo 2B models on the cloud
  2. The generated videos by text appears as a "pond5" watermark

Expected behavior / 期待表现

I hope the officials can explain and provide solutions as much as possible.

zRzRzRzRzRzRzR commented 2 months ago

I believe this should be the 2B model, not the 5B model. The 2B model has a slight issue during the SFT stage, which causes it to have a pond5 watermark under certain prompts. We haven't maintained this version of the model much. Our focus of research is on the 5B model, which does not have this issue. Currently, the issue with the 2B model cannot be resolved.

TemporalLabsLLC-SOL commented 1 month ago

I can confirm 5b does do it rarely. I'm cataloging out every gs and steps combination possible on my hardware to see if there's any pattern or reason to it.

My Temporal Prompt Engine allows full localized batch processing including video sorting and watermarking so I'll be sharing that data when I have it compiled.

Once I can get my founderpass going I'll get my aws station access back and I can pick up the pace.

Hopefully we can just avoid certain maths.