qguopku comments

Results 29 comments of


                                            qguopku

When can your training weights be shared?

Hi, you can use the latest train_hack.py, I will release it once I have trained a robust enough model.

sample = sample + latent_pose

hack是我自己实现的，非hack的是原文的（非常难控制pose，需要大规模数据+overfit）

sample = sample + latent_pose

> > hack是我自己实现的，非hack的是原文的（非常难控制pose，需要大规模数据+overfit） > > 原来如此，感谢大佬，数据量小的话hack的效果更好些是吗？对，即使数据量大，我也建议使用hack

Is a device with 2*4090(24G) able to do the trainning?

It is frustrating to note that the current training code requires a minimum of 40G VRAM for the A100.

Is unet trained？

yes

Is unet trained？

> @guoqincode hello sir, that means after stage1 training, the unet cannot generate content based on the prompt word? if i set unet.requires_grad_(False) and unet.eval(), will unet still gain the...

Hello! I would like to remind you of your fast_ env.sh file name is incorrect

Thank you!

有办法降低一下显存占用吗

I'm sorry, but I found that tricks (FP16, BF16, etc.) that reduce the memory will reduce the performance of the model (NAN loss often occurs), so 80GB VRAM is best...

pre-made motion cap movement packs

You can use: https://github.com/IDEA-Research/DWPose

mat1 and mat2 shapes cannot be multiplied (4x768 and 1280x320)

Hello, have you changed the code? I use the same code that is trainable.