使用自定义目标人物生成的视频嘴形异常 - Abnormal Avatar Mouth Movement

[Qifeng-Wu99]

感谢作者的精彩工作。

目前我在尝试用自定义目标人物视频来训练postnet和radnerf，但推理的结果令人不满意，生成的视频中人物嘴形没有跟随语音，处于一种似张非张的状态。

目前我用adobe premiere pro剪辑目标人物视频用于训练，参数为30fps，大小512*512，mp4格式。

预处理及训练过程严格按照作者在说明文档中的进行。

请问：

1. 对于自定义目标人物的视频，有无更加严格具体的要求
2. 能否对于调试（debug）给出指导意见

谢谢

[Qifeng-Wu99]

I really appreciate the fantastic job done here by the authors.

I am currently using a customized video to train postnet and radnerf. However, the results by the time of inference is unsatisfactory. That is the avatar's mouth is not following the audio, alternating between shut and slightly open.

The video of my target identity is created with adobe premiere pro, I reduced the frame rate from 60fps to 30 fps, cropped the frames to 512*512 and saved it as mp4 file.

I believe I follow the preprocessing and tranining guidance given by the author.

I wonder if there are extra requirements of the video of target identity and I would appreciate any advice on debugging.

Thanks in advance.

[jinqiupeter]

The video to use MUST be in 25 fps

[Qifeng-Wu99]

The video to use MUST be in 25 fps Thanks for your reply, I will update you if things getting better.

[Qifeng-Wu99]

The video to use MUST be in 25 fps

Thank you so much for your help, this tip proved to be working very well.