LeslieZhao

Hi @shiminxuan003 1024无法预载256的模型的，训练的前提是stylegan2的模型是可以支持生成真实人的，如果你想在1024上训练，建议你找到或者自己基于ffhq训练一个stylegan2的模型。 至于'NaN or Inf found in input tensor.' 我不确定你的stylegan2预载的模型是否有什么问题，或者你的输入输出，学习率设置上，你可以使用pdb进行断点debug，查看输入输出是否是符合预期的。 你可以使用tensorboard查看中间结果，理论上刚开始训练的时候xg和xt的相似度是很大的。 期待你有一个好的结果！

Hi @darkasevgen ! I didn't use early stop. It is just a template. If you want to use it, just modify the evalution function and [get_loss_from_val](https://github.com/LeslieZhoa/DCT-NET.Pytorch/blob/main/trainer/ModelTrainer.py#L182). As described by Readme，when...

self.args.scratch_d_path is here https://github.com/LeslieZhoa/GFPGAN-1024/releases/download/untagged-7fadd91b517da1810e55/d.pth, I will fix the link later

I think you can train a good model yourself according to the tutorial, so I don't have any plans for an open source model at the moment

> > > Я думаю, что вы можете самостоятельно обучить хорошую модель по учебнику, поэтому на данный момент у меня нет планов на модель с открытым исходным кодом. > >...

You can follow https://github.com/LeslieZhoa/LVT and set it to you sys.path like [Here](https://github.com/LeslieZhoa/DCT-NET.Pytorch/blob/1f7a8adb13f8603db0ff8303560fcddb7aeb7033/utils/get_face_expression.py#L91)

fix it, https://github.com/LeslieZhoa/GFPGAN-1024/releases/download/v0.1/d.pth

Thanks your reply. You can also change [here](https://github.com/LeslieZhoa/HeadSwap/blob/main/process/process_func.py#L19) to face_alignment.LandmarksType.THREE_D

It is not bad in simple scene with small movement amplitude. You need smooth the keypoints

It does have a lot of room for improvement.