Zhenhui Ye

Results 41 comments of Zhenhui Ye

Hi, please check the loss curves of lm3d_radnerf_torso, I suspect that the training of lm3d_radnerf_torso is failed.

Hi, the PSNR curve seems normal, can you refer to the validation results generated during training, it should be something like this: `checkpoints/May/lm3d_radnerf_torso/validation_results/validation_250000/images/frame_5573.png`

Hi, as the validation_results could obtain reasonable results (a predicted torso), we can confirm that the training process is performed normally. So I suspect it should be something wrong with...

你好,大概率是因为postnet没有被很好的训练导致的。为了验证这点,您可以使用utils.visualization.lm_visualizer.py来可视化postnet生成的landmark文件,然后观察其是否与语音对齐。因为GeneFace的postnet的训练依赖GAN loss,导致其训练不太稳定,得到效果好的ckpt需要一定的技巧,具体可以看一下我们在train_models.md里面提到的技巧。如果您愿意等的话,可以等一波我们会在8月份发布的GeneFace++代码(会作为本仓库的一个release),在其中postnet的训练难度得到大大简化。

你好,GeneFace和GeneFace++属于person-specific的方法,即在May视频上训练的模型只能用于生成May。如果需要用单张图片驱动,可以尝试一下SadTalker等单图驱动的方法 -----原始邮件----- 发件人:DAIPING ***@***.***> 发送时间:2023-07-28 17:33:24 (星期五) 收件人: yerfor/GeneFace ***@***.***> 抄送: "Zhenhui Ye" ***@***.***>, Comment ***@***.***> 主题: Re: [yerfor/GeneFace] 请教一下中文说话人生成嘴部同步问题 (Issue #152) 您好,请教一下。 当前我将代码中的Demo May已经能够跑起来了,如果我想使用【一段音频】+【1张图片】+【使用May训练出来的模型】来生成最终结果的视频,这样的操作是可以的吗?如果是可以的,我该如何操作? — Reply to this email...

Hi, 或许你可以用别的python环境的tensorboard来显示log。或者重新安装geneface环境。

你好,之前没有遇到显存不足的bug。可否提供更多细节,比如可以设断点debug一下各个阶段,load 模型前后显存占用的变化。会不会是别的进程占用了显存?

Hi, maybe this will work: ``` CUDA_VISIBLE_DEVICES=0,1,2 python tasks/run.py --config=fs.yaml --exp_name=fs_test_demo --reset ```

It seems like a error caused by index out of bounds. Can you provide more details? Since the code should have convert the audio to 16k and video to 25...