Tao Liu

I am also searching for this answer. I guess that the continuous space is more smooth and controllable than the VQ-based space. ===update===== After reviewing the paper, here are my...

@Wangman1 I have trained on HTDF and the tendency is similar to you: vgg loss is high (~ 50) and g loss (~7) is larger than d loss(~0.0x). Although I...

@Wangman1 This is my training loss. I am fixing on it...

@Wangman1 可以的。这也是我觉得奇怪的地方，loss 看着不太对，但是可以驱动成功。

@Wangman1 我是先载入作者 release 的 vox 模型，然后再去 hdtf 训练的，作者应该用的是 vox1 数据集，这个数据集大概有 400 多万张图片，你用的是多少？如果太少的话确实有可能训练不成功。训练是 8 卡 A10， 10w 次迭代大概用了 1 天时间，我这里是机械硬盘，GPU 利用率还没有打满，速度上还有提升空间。

@Wangman1 This is the loss trained on vox1 from scratch by myself , maybe it will help.

@wyhsirius Hello, I would like to inquire about the final L1 loss in your training regimen. I have trained on Vox1 with your training parameters and L1 loss achieved 0.08...

@johndpope , great job. However, I have some suggestions after watching some videos from CelebV-HQ. You may manually filter them to ensure high resolution and remove bad cases, such as...

Same Issue

+ 1