face2comics by Sxela (Alex Spirin)
This is a paired face to comics dataset, which can be used to train pix2pix or similar networks.
This dataset contains a lot of crappy nightmare-fuelish samples, which tend to be useful for full image to comic convertion, as they teach models how to trace something if it's not a face.
Update: v2.0.0 dataset added!
v2.0.0 dataset
Darkish red style, 1024x1024, 10000 pairs (total 20k images) Download faces from github (~1Gb) Download comics from github (~1Gb)
2x2 dataset sample:
10x10 dataet sample:
An inference sample of a fastai unet trained on this very dataset :
You can train your own unet by using their awesome notebook here
v1.0.0 dataset
Dark style, 512x512, 10000 pairs (total 20k images) Download from github (~700Mb)
2x2 dataset sample:
10x10 dataet sample:
An inference sample of a fastai unet trained on this very dataset :
You can train your own unet by using their awesome notebook here
Recognition
justinpinkney for his layer blending idea, repo and tutorials https://github.com/justinpinkney/stylegan2
NVlabs for their Stylegan2, of course, though I didn't use their original code https://github.com/NVlabs/stylegan2
fastai for their awesome framework and simple GAN training pipeline https://fast.ai