Comment by ansk | Hacker News Reader

Comment by ansk | Hacker News Reader

Image generative models have made significant progress over the last few years, but an often-overlooked confounding variable is that during this period, the predominant benchmarking task used by researchers (specifically those working on large-scale image synthesis) switched from unconditional image synthesis to text-to-image synthesis. Text-to-image synthesis is a fundamentally easier task as the text-conditioned distribution is much more predictable than the unconditional counterpart. Additionally, text-to-image plays to the strengths of diffusion models, which benefit from having a conditioning signal to guide the sampling process (via classifier(-free) guidance). As someone who is observing this progress from an adjacent field, it has been hard to tease out how much of this seemingly abrupt progress can be attributed to the fact that the preferred benchmarks and datasets have been shifting behind the scenes. I suspect that if these variables were well controlled, progress over the last few years would look more like the steady, incremental improvements that occurred in the field in the years prior.

Maybe I'm misreading the tone of your comment - but is that a bad thing? Reframing the approach to a problem, to state-of-the-art results; that strikes me as nothing if not progress. Don't mean to come across as negative here -- I just don't see why anyone would control those variables.

You list reasons why text-to-image is easier than unconditional image gen - but isn't that the point? They're both under the umbrella of image generation, which is ultimately pushed forward.

You can give a blank prompt to stable diffusion and it will give you reasonable images https://www.reddit.com/r/StableDiffusion/comments/xu0dt6/no_...

The good old GAN models were trained on single domain dataset or on Imagenet, and these datasets only encode a small set of the real distribution, make them easier to fit, also I don't think GANs are flexible enough to be effectively used in text-to-image generation, so I don't think it's an easier task, it was a more difficult task and it required the development of autoregressive and diffusion model to actually make it feasible.

BigGAN trained on JFT-300M and it worked fine. In fact, JFT-300M was easier than ImageNet. In their own words:

"Interestingly, unlike models trained on ImageNet, where training tends to collapse without heavy regularization, the models trained on JFT-300M remain stable over many hundreds of thousands of iterations. This suggests that moving beyond ImageNet to larger datasets may partially alleviate GAN stability issues."

This suggests GAN will work even better on, say, LAION-5B. It's just that nobody tried.

"GANs are not flexible enough to be effectively used in text-to-image generation" is an absurd statement. GANs have latents, there is no reason why just feeding CLIP to GANs shouldn't work. Again, nobody tried because most GAN works preceded CLIP.

We should try GANs again. It was abandoned without any evidence whatsoever that diffusion model etc is required. GANs have genuine advantages over diffusion models, like much faster sampling.

"It was abandoned without any evidence whatsoever that diffusion model etc is required.", diffusion models beats GAN models in quality on all datasets it was tested on, I suggest to read (again probably): "Diffusion Models Beat GANs on Image Synthesis" [1], the difference in quality is astounding, also the fact that GANs seems to work well with small one domain dataset, then they became incrementally more difficult to train with Imagenet and then it became easier again with 300M images only demonstrates how sensible these models are to hyperparameters and unstable they are to train (some distributions are better than others), these are the two main reasons why researchers have switched from GAN to Diffusion and autoregressive models, I like your attitude towards protecting the legacy of GANs but I don't put too much faith in them if not for autoencoders and upscalers (although some NN such as SwinIR have completely abandoned the adversarial loss). I bet the future for generative models will be distilled diffusion models for fast sampling. Btw you comment has been clearly influenced by an article [2] from Gwern blog but you didn't link it.

[1]: https://arxiv.org/abs/2105.05233 [2]: https://www.gwern.net/GANs

Reply on news.ycombinator.com