镜头前后跳

[huangxin168]

非常优秀的项目！感谢分享！ 我按照安装说明配置好linux环境后，测试了一下， python inference.py -s assets/examples/source/s0.jpg -d pose0.mp4 --flag_crop_driving_video 发现生成的结果镜头很不稳定，前后跳跃，不清楚什么原因导致的问题。 请大佬抽空看看。 谢谢！

https://github.com/user-attachments/assets/c43cc26c-a28e-4092-bea6-9604d4849bf8

[cleardusk]

@huangxin168 感谢反馈，可否把原始驱动视频贴上来 : )

Thank you for the feedback. Could you please share the original driving video? :)

[huangxin168]

@huangxin168 感谢反馈，可否把原始驱动视频贴上来 : )

Thank you for the feedback. Could you please share the original driving video? :)

https://github.com/user-attachments/assets/190574d2-81ad-4f18-9490-9e85bb25007a

[cleardusk]

@huangxin168 感谢提供，我们正在定位中。

Thank you for providing it. We are currently debuging the issue.​

[cleardusk]

It seems that the scale is not stable, especially when the input texture is warped (e.g., by D-ID). We will continue to work on optimizing this issue. @huangxin168

You can try to input a realistic driven video, similar to the example videos!

[huangxin168]

It seems that the scale is not stable, especially when the input texture is warped (e.g., by D-ID). We will continue to work on optimizing this issue. @huangxin168

You can try to input a realistic driven video, similar to the example videos! Thanks for your debugging. Can we fix the scale when inferencing?

[UltraClr]

我也有类似的问题，面部抖动，我是采用论文中提到的音频驱动的方式生成的。请问作者有什么解决方式吗

https://github.com/user-attachments/assets/9796339b-81ef-4193-8946-f33e35a25e6e

https://github.com/user-attachments/assets/781e7c92-aa27-4b2c-94d1-423b3620c934

[aihacker111]

It's related about the frame image paste list , seem like when after combine audio to video the frame is not continuously , I'm working on the new feature for multiple motion face in multiples face image, my result is better when not have audio , and have audio, the result is very confused

[FacePoluke]

Modify the code at live_portrait_pipeline.py#L284 to : scale_new = x_s_info['scale'] can solve the problem https://github.com/KwaiVGI/LivePortrait/blob/d654a014da85e4b45d17b0c2016acc843a392149/src/live_portrait_pipeline.py#L284

[huangxin168]

Modify the code at live_portrait_pipeline.py#L284 to : scale_new = x_s_info['scale'] can solve the problem

https://github.com/KwaiVGI/LivePortrait/blob/d654a014da85e4b45d17b0c2016acc843a392149/src/live_portrait_pipeline.py#L284

https://github.com/user-attachments/assets/5fb2be08-00f3-4ddb-b9ee-402dfa57408a Thanks for your help, I tried and got new problem when eye-blinking.

[wangqiang9]

Modify the code at live_portrait_pipeline.py#L284 to : scale_new = x_s_info['scale'] can solve the problem

https://github.com/KwaiVGI/LivePortrait/blob/d654a014da85e4b45d17b0c2016acc843a392149/src/live_portrait_pipeline.py#L284

helpful!

[markson14]

我也有类似的问题，面部抖动，我是采用论文中提到的音频驱动的方式生成的。请问作者有什么解决方式吗

s6--test_FaceTalk_170809_00138_TA_condition_FaceTalk_170913_03279_TA.mp4 test_FaceTalk_170809_00138_TA_condition_FaceTalk_170913_03279_TA.mp4

这真的太棒了，请问你这是如何使用音频驱动的方式？是直接修改代码就可以吗？还是需要自己搭建一个训练框架来实现这个audio-driven模块？谢谢。

[bageers0]

似乎没有用，效果依旧差，跳来跳去的 @FacePoluke https://github.com/user-attachments/assets/37760ae5-00aa-4bfe-9880-775fc704dcf9