3 ms·

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 ImageNe

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.

GaggiX 4y ago
"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 https://arxiv.org/abs/2105.05233 [2]: https://www.gwern.net/GANs https://www.gwern.net/GANs

GaggiX 4y ago

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