GigaGAN: Large-Scale GAN for Text-to-Image Synthesis | Hacker News Reader

GigaGAN: Large-Scale GAN for Text-to-Image Synthesis

mingukkang.github.io

37 pointsGaggiX10 comments

>We present our 1B-parameter GigaGAN, achieving lower FID than Stable Diffusion v1.5, DALL·E 2, and Parti-750M. It generates 512px outputs at 0.13s, orders of magnitude faster than diffusion and autoregressive models, and inherits the disentangled, continuous, and controllable latent space of GANs.

Nice to see GANs making a resurgence, as they are much faster and more controllable. Hands seem to still be messed up though. Are hands just a dataset issue or is there something really hard about it? My first thought is that faces would seem to be more difficult and that has already been mostly solved.

I heard it once explained as "most fingers have a finger either side" issue. That said, there certainly are SD models that are quite solid in this regard, like deliberate. I don't know much about GANs but if the problem is shared w diffusion, perhaps the solution is as well and we just need better training.

Hands compare to faces have a lot of limbs, they can take many different poses that makes it difficult for the model to learn them, also these models are usually not good with "counting", an hand has 5 fingers but you can easily expect that a model will create an hand with 4, 6, 7 fingers.

Hmm. Little discussion of artistic styles (a few examples of using a particular medium like oil, or a movement like expressionism, or referencing a time period like Age of Sail), and no examples of a specific artist or work, not even the ubiquitous Claude Monet or Starry Night.

I find it a bit hard to believe the model is completely incapable of rendering these prompts, unless the training data was strictly filtered to exclude artist names and their styles as much as possible.

I've got 2 questions:

1) Can I run this, say, on Colab? I see no code.

2) Any interactive open source tool to visually explore the latent space of a model’s representations for any given prompt, or for faces, buildings, etc.? Like they exist for fonts [1].

[1] https://pair-code.github.io/font-explorer/

I don't think that Adobe is going to release anything unfortunately.

Can someone tells me what is the point that so many researchers are spending time generating these weird images that seem neither artistic nor interesting?

So that even the dullest of HN users will find them interesting and artistic.

Link to the paper: https://arxiv.org/abs/2303.05511

Very cool, I'd love to see a result dump of random samples

Reply on news.ycombinator.com