生成质量下降严重

[mingzhe11]

video_gen_tbiXjKuayn9ISSxCkhAw1764293501903_edit_compare.mp4

video_gen_qswzymcGC8pQoJDfELU41767059253494_edit_compare.mp4

video_gen_o625IGTqaSpr1iMwcbYH1764293574663_edit_compare.mp4

video_gen_D_NRdoZMyzGce5L1r5x51767059333911_edit_compare.mp4

发现人物一致性不是很好，生成质量下降地比较严重（眼睛、嘴巴等）

video_gen_8u3lQA0CwQ779aTU053i1764294105771_edit_compare.mp4

20251202-173109_edit_compare.mp4

还有指令遵循等问题

请问我如何设置可以提升编辑质量，我用的就是diffuser_demo.py。

[linyq17]

Thanks for your interest in our work!

There are two factors that significantly impact the results.

• First, the prompt and training data need to be aligned as closely as possible — you can refer to the prompt tips from link and the OpenVE Bench for guidance.
• Second, we found that different random seeds also have a substantial effect on the final outcomes. You can add a multi-seed sweep in the code to select the best outcome

[xuhang07]

Hello, thank you very much for open-sourcing such amazing work! I have a question regarding the generation quality. We noticed that the results are excellent when seed=0, but the quality drops significantly when switching to other seeds. Additionally, the CFG scale seems to have a minimal impact on the results. I was wondering if the released weights are distilled, or if there were specific training settings that led to this phenomenon?

[linyq17]

Thank you for the detailed observations!

Regarding seed sensitivity: The model is indeed sensitive to train-test alignment, which is why seed=0 is our recommended default setting. We observed that this seed produces the most consistent results aligned with our training configuration.

Regarding CFG scale: The released weights do not support classifier-free guidance (CFG). To improve convergence efficiency during training, we did not include unconditional training. As a result, varying the CFG scale does not work with our model.

[mingzhe11]

Thanks for your interest in our work!

There are two factors that significantly impact the results.

• First, the prompt and training data need to be aligned as closely as possible — you can refer to the prompt tips from link and the OpenVE Bench for guidance.
• Second, we found that different random seeds also have a substantial effect on the final outcomes. You can add a multi-seed sweep in the code to select the best outcome

好的，感谢你的回复。我去多测测