stylegan3 VS StyleGAN3-CLIP-notebooks

StyleGAN3 by Nvidia Open Source Software »

StyleGAN by Nvidia (Open Source) - GitHub » StyleGAN on GitHUB

Not unless you are Nvidia Corporation: https://github.com/NVlabs/stylegan3/blob/main/LICENSE.txt

I found the follow script but unsure of how to utilize it in this context. https://github.com/NVlabs/stylegan3/blob/main/torch_utils/persistence.py

full dataset.py file that I'm changing is here: https://github.com/NVlabs/stylegan3/blob/main/training/dataset.py

The neural network on which the training was carried out - https://github.com/NVlabs/stylegan3

I trained a Generative Adversarial Network stylegan2-ada model with NVIDIA’s StyleGAN3 algorithm with a RTX 3090 GPU for a few nights. This is an interpolation video of the generator model’s random-walk datapoints divided into 4 different windows. Initially I web crawled some 2000 images of any Formidable images and cropped them with nagadomi’s lbpcascade_animeface anime face detector, with a setting that I attempted to also include her assets in the image. Previously I have done by transfer-learning from Gwern’s ThisWaifuDoesNotExist, which only included heads of Emilia from Re:Zero, which was quite good. This time I wanted to see if the model can also handle having something more than just a head. Having Formidable’s chest also in the image made some angles perform pretty bad, as there are as many ways of making anatomy as there are artists. Because of this, I removed all swimsuit and party skin images, as making her features was hard enough with her default skin, making the final dataset size some 1500 images. In the end, I’m pretty satisfied with the results, but I could prune the dataset even more and crop the images more homogenously as well as try a bit different hyperparameters (most importantly gamma) and stylegan3-t. However, I want to move into trying out Stable Diffusion model, so I will wrap this project up at least for now and post this. There is a psi hyperparameter used in this video generation, that determines how “creative” the generator might be, i.e. how far from an optimal statistical distribution it can go at any given datapoint (video time in this case). With psi=0 you have almost static video, and with psi=1 wildly varying results of which half aren’t even recognizable, and some are really good. I settled for 0.65, which I think has some nice variety with a reasonable amount of bad morphs.

Refer to the training configurations to choose the correct values for your hardware, config and dataset. Reducing the --batch value typically warrants increasing --gamma and/or lowering the D and G learning rates (--dlr and --glr). Think of --gamma as the “randomness” or “creativity” of the model.

Heck to train StyleGAN, an AI image style transfer model, requires 1-8 GPU’s with a minimum of 12GB of VRAM to train. Using a single Tesla V100, which is about £6-10k, it still can take upwards of one minute to process 1000 256x256 images.

Actors were generated with StyleGAN3+CLIP, text was written by latent-diffusion, names were generated using this site.

Since it's a work in progress, I'll also share this repo where I've been updating the notebook.