search

lucidrains / stylegan2-pytorch

Simplest working implementation of Stylegan2, state of the art generative adversarial network, in Pytorch. Enabling everyone to experience disentanglement
https://thispersondoesnotexist.com
MIT License
3.71k stars 587 forks source link

Mode Collapse! #223

Open gamladz opened 3 years ago

gamladz commented 3 years ago

Hello,

We're trying to use PyTorch on our furniture dataset but keep encountering mode collapse. Have attached the 5th and produced images. I've read through other folks strategies with the same issues and tried using the following settings.

5 8

stylegan2_pytorch --data PNG --multi-gpus --attn-layers [1,2] --batch-size 8 --gradient-accumulate-every 16

Any ideas?

lucidrains commented 3 years ago

@gamladz more data and higher augmentation probability!

thusinh1969 commented 3 years ago

StyleGAN current plus ADA whatever can only work with very portrait and focused-simple to detect dataset (faces). Do NOT expect any good results at all with other perplexed dataset. It is the capacity of network and its current design will NOT do the job.

have been trying a lot on models, twisted and trained for whatever resolution, GPU etc. it simply can not learn anything good but faces !!!

Not interesting at all this kind of “demo for face only models”.

Steve