4 ms·

I am looking for someone who can help me create (or help me get started to create) a GAN that generates H.R. Giger style pictures. Anyone up for helping me out

I am looking for someone who can help me create (or help me get started to create) a GAN that generates H.R. Giger style pictures. Anyone up for helping me out ?

gwern 8y ago
Use StyleGAN. The speed and quality of its results surpass any GAN I've ever used, and I've used dozens of different implementations of various architectures and tweaked them over the past 3 years. If you want a demo of what it can do and the original Nvidia paper/video isn't impressive enough, check out the anime faces I've been generating with just a few days of training on my 2 GPUs: https://twitter.com/gwern/status/1093701790971953152 https://twitter.com/gwern/status/1093701790971953152 For H.R. Giger pictures, I suspect the problem is that you simply don't have enough. StyleGAN is powerful enough to memorize datasets of under a few thousand images. What I would suggest is that you might be able to take the Nvidia 1024px face model or cat model and retrain the model with your artwork; that saves you sample size, and to the extent that faces or cats remain in your final model, well, perhaps that will just make it creepier or hilarious, in the vein of 'dogslugs'. (This is what I have been doing to make individual-character anime-face StyleGAN models: few characters have so many faces as to be able to train from scratch, so I've trained a StyleGAN on the entire face dataset, and then I simply retrain it for an hour or two on just that individual character's faces to finetune it.) Alternately, you could try to borrow strength from an illustration dataset somewhat similar, maybe a Western art dataset like WikiArt? (I forget if WikiArt is the dataset OP mentions Mario training his GANs on.) Train a StyleGAN on that and then do transfer to Giger as above. Finally, you could get more invasive and try to weaken StyleGAN enough to make memorization more difficult when training from scratch. Shrink the 8x512 FC layer stack and remove filters from the convolutions, for example. It will be a lot weaker but it'll generalize better within the samples it's still able to generate. Train faster/smaller too.
- raymond_goo 8y ago
  Wow, thanks for the thorough reply!

I am looking for someone who can help me create (or help me get started to create) a GAN that generates H.R. Giger style pictures. Anyone up for helping me out ?

gwern 8y ago

Use StyleGAN. The speed and quality of its results surpass any GAN I've ever used, and I've used dozens of different implementations of various architectures and tweaked them over the past 3 years. If you want a demo of what it can do and the original Nvidia paper/video isn't impressive enough, check out the anime faces I've been generating with just a few days of training on my 2 GPUs: https://twitter.com/gwern/status/1093701790971953152 https://twitter.com/gwern/status/1093701790971953152 For H.R. Giger pictures, I suspect the problem is that you simply don't have enough. StyleGAN is powerful enough to memorize datasets of under a few thousand images. What I would suggest is that you might be able to take the Nvidia 1024px face model or cat model and retrain the model with your artwork; that saves you sample size, and to the extent that faces or cats remain in your final model, well, perhaps that will just make it creepier or hilarious, in the vein of 'dogslugs'. (This is what I have been doing to make individual-character anime-face StyleGAN models: few characters have so many faces as to be able to train from scratch, so I've trained a StyleGAN on the entire face dataset, and then I simply retrain it for an hour or two on just that individual character's faces to finetune it.) Alternately, you could try to borrow strength from an illustration dataset somewhat similar, maybe a Western art dataset like WikiArt? (I forget if WikiArt is the dataset OP mentions Mario training his GANs on.) Train a StyleGAN on that and then do transfer to Giger as above. Finally, you could get more invasive and try to weaken StyleGAN enough to make memorization more difficult when training from scratch. Shrink the 8x512 FC layer stack and remove filters from the convolutions, for example. It will be a lot weaker but it'll generalize better within the samples it's still able to generate. Train faster/smaller too.

raymond_goo 8y ago
Wow, thanks for the thorough reply!