akanimax / BMSG-GAN

[MSG-GAN] Any body can GAN! Highly stable and robust architecture. Requires little to no hyperparameter tuning. Pytorch Implementation
MIT License
630 stars 105 forks source link

Msggan quality results #50

Open Satoszi opened 1 year ago

Satoszi commented 1 year ago

Hello!

I am using msggan, but I get worse results than in the paper.

I tried celebHQ whole dataset (100k samples), with resolution 64x64 and 128x,128. For 64x64 I have done ~250 epochs and for 128x128 ~ 47epochs.

Results are not bad (much better than DCGAN), but they have worse texture quality than in msggan paper, and only 30% samples looks realistic. 70% samples look like monster (weird faces, weird artifacts etc).

I am using 3060ti batch size 12 Rest of hyperparameters are default like in repo (LR = 0.003, latent_size = 512, loss_function = relativistic-hinge, flip-augment=True)

Any ideas? Can anyone share results?

image

I am attaching my 64x64 results (every epoch is 100k samples, so 251 epoch is 25mln real samples)

image

and 47 epoch for 256x256