Why Bigger Isn’t Always Better with GANs and AI Art | Hacker News Reader

Why Bigger Isn’t Always Better with GANs and AI Art

artnome.com

102 pointsArtnome6 comments

Great article! Thanks for the writeup, resources and introducing me to Helena Sarin.

I've been wanting to get into GANs but I've really been turned off by the samey-ness of the work I've been seeing. I'm afraid I'd spend a ton of time getting everything set up only to end up creating more 512x512 jpgs that look like everything else everyone else is making. Helena's work makes me want to give it another shot though.

To me, it really feels like AI-generated artwork is in that hype phase you see with any new tech used in art. Would that work from Obvious sold for nearly half a million dollars if it wasn't a GAN-generated work? Probably not. I'm looking forward to the day an AI work gets that much attention because it's compelling all on its own, not just because of how it was made.

For comparison, here's a web app that lets you play with BigGAN:

https://ganbreeder.app/

appears to be down?

One way forward would be retraining the official BigGAN models. You may not have enough compute to train BigGAN from scratch, but you may have enough to finetune/retrain it on a new dataset to get the best of both worlds.

There's an attempt at implementing BigGAN in PyTorch (https://github.com/AaronLeong/BigGAN-pytorch) - anyone know how to load & retrain a TF model in that (https://github.com/AaronLeong/BigGAN-pytorch/issues/2)? :)

Main point: bigger GANs like BIGGAN don't let you train with your own images because they're so expensive to train. So everyone is creating art from the same pretrained model (and it'll presumably all look somewhat similar)

> bigger GANs like BIGGAN don't let you train with your own images because they're so expensive to train.

I don't think we know that. They're super-expensive to train from scratch, but there have been few experiments investigating GAN transfer learning. In my own very simple personal experiments, when I trained a ProGAN on one anime character's faces and then retrained it on a second anime character's faces (Asuka->Holo), it adapted very quickly and successfully; so I suspect that the released BigGAN models might be retrainable with a tiny fraction of the resources. Training on your own photos should be doable.

The real reason that everyone is using the BigGAN models and not retraining them is because, well, there's no way to retrain them! The BigGAN source wasn't released, and the only implementation so far doesn't support loading them and retraining. So you can't do it, no matter how much you want to, unless you are so good at TF/PyTorch that you can program it yourself. Which apparently no one has so far.

Reply on news.ycombinator.com