4 ms·

>I'm sorry, then I don't know what you're saying I mean it was pretty simple, the quality is not really close, also StyleGAN2 is not conditioned on text (becau

by GaggiX 3y ago

>I'm sorry, then I don't know what you're saying I mean it was pretty simple, the quality is not really close, also StyleGAN2 is not conditioned on text (because you talked about this model in first comment). In the future there could be one that is competitive but not today. GigaGAN the best GAN model by far is not open source and it's not competitive yet because even the images cherry picked for the paper and project paper do not look that coherent, the FID is relatively low because the inception v3 model used to calculate the FID doesn't care that much about global coherency and more like texturing but if the FID was calculated using DINO v2 (like some recent paper do) instead of inception v3 it will really show the gap between GAN models and diffusion models today.

godelski 3y ago
Look, every single message I've said that you're trading quality for speed. That does in fact mean GANs are worse. I'm not sure why you think I have said anything short of that, I've explicitly agreed with you that diffusion produces higher quality images, and I'm not interested in repeating myself any further. We're talking about a fucking raspberry pi, it is definitely reasonable to want to __trade quality for speed__ when you're working on a tiny computer. I'm glad you're aware that FID has limits. For some reason this is uncommon. But there are more limits than the backbone classifier. Yes, DINOv2 helps (so does CLIP-FID and clean FID (which is backbone independent. I feel I need to explain because we're having difficulties communicating and going to their code and seeing they use inception isn't going to mean anything to be because that's not what that work is about)), but fundamentally the difference between two normalization layers of a classifier is not actually measurement of image quality. It correlates, but these also are fundamentally about the distributional nature of the outputs. There actually is no method that does particularly well but we're over here doing our best. Just got to be aware of the limits of your metrics because (as is the theme of our entire conversation), context matters. Here's a paper you may find interesting https://arxiv.org/abs/2306.04675 https://arxiv.org/abs/2306.04675
- GaggiX 3y ago
  Running SD v1.5 using one single step is faster than sampling with GigaGAN and you can achieve better coherency in my opinion: https://arxiv.org/abs/2311.09257 https://arxiv.org/abs/2311.09257 Then it does make more sense to run this on a raspberry pi.
  - godelski 3y ago
    Yeah I'll agree to that. But neither UFOGen nor GigaGAN have released models (I've admitted to being wrong about GigaGAN's checkpoints. In fairness, who the fuck releases 6GBs of generated images? And removes the issues tab from GitHub? Can we agree that's sketchy as fuck?[0]). LAFITE is the only of the 3 with released checkpoints but it's definitely not to the same quality (Nvidia backed down to place it in products and imo a bad move). But it is 75M params compared to the 1B of UFOGen and GGAN. I was able to get a bit better images by some prompt engineering but yeah, classic paper painting a much better picture of their model than it actually is (pun intended). But then again, I got shitty images the first time I used SD so YMMV. There's probably better works out there but honestly my focus is elsewhere and no one's got time to keep on top of and try everything coming out. I'm literally just saying that you can trade quality for inference. Do you disagree? Because I'm not sure why you think I think GANs are better for quality, as I've said the opposite many times. Why are you hyper-focused on quality and thinking I've said GANs do better. This is why we're talking past one another, because you're attributing assumptions to me that I ('m doing my fucking best to clarify that I) don't have. If you think I think GANs produce higher quality images, I assure you that this is from your imagination as I've never stated such and it is not an opinion that I have. Sorry, I said I wouldn't say this again, so last time for real. [0] There's too much sketchy shit going on in ML research right now and honestly, that's why I've been more passionate about trying to get people to think harder and about context. Again, this is about context. (I also really hate this railroading as it stifles community innovation and sweeps important problems under the rug by saying to just rely on large companies for checkpoints. The community doing so much around Stable Diffusion is awesome, but I want to see that around lots of works because there's a lot that can be accelerated by even a hundredth of this community effort)

by GaggiX 3y ago

godelski 3y ago

Look, every single message I've said that you're trading quality for speed. That does in fact mean GANs are worse. I'm not sure why you think I have said anything short of that, I've explicitly agreed with you that diffusion produces higher quality images, and I'm not interested in repeating myself any further. We're talking about a fucking raspberry pi, it is definitely reasonable to want to __trade quality for speed__ when you're working on a tiny computer. I'm glad you're aware that FID has limits. For some reason this is uncommon. But there are more limits than the backbone classifier. Yes, DINOv2 helps (so does CLIP-FID and clean FID (which is backbone independent. I feel I need to explain because we're having difficulties communicating and going to their code and seeing they use inception isn't going to mean anything to be because that's not what that work is about)), but fundamentally the difference between two normalization layers of a classifier is not actually measurement of image quality. It correlates, but these also are fundamentally about the distributional nature of the outputs. There actually is no method that does particularly well but we're over here doing our best. Just got to be aware of the limits of your metrics because (as is the theme of our entire conversation), context matters. Here's a paper you may find interesting https://arxiv.org/abs/2306.04675 https://arxiv.org/abs/2306.04675

GaggiX 3y ago
Running SD v1.5 using one single step is faster than sampling with GigaGAN and you can achieve better coherency in my opinion: https://arxiv.org/abs/2311.09257 https://arxiv.org/abs/2311.09257 Then it does make more sense to run this on a raspberry pi.
- godelski 3y ago
  Yeah I'll agree to that. But neither UFOGen nor GigaGAN have released models (I've admitted to being wrong about GigaGAN's checkpoints. In fairness, who the fuck releases 6GBs of generated images? And removes the issues tab from GitHub? Can we agree that's sketchy as fuck?[0]). LAFITE is the only of the 3 with released checkpoints but it's definitely not to the same quality (Nvidia backed down to place it in products and imo a bad move). But it is 75M params compared to the 1B of UFOGen and GGAN. I was able to get a bit better images by some prompt engineering but yeah, classic paper painting a much better picture of their model than it actually is (pun intended). But then again, I got shitty images the first time I used SD so YMMV. There's probably better works out there but honestly my focus is elsewhere and no one's got time to keep on top of and try everything coming out. I'm literally just saying that you can trade quality for inference. Do you disagree? Because I'm not sure why you think I think GANs are better for quality, as I've said the opposite many times. Why are you hyper-focused on quality and thinking I've said GANs do better. This is why we're talking past one another, because you're attributing assumptions to me that I ('m doing my fucking best to clarify that I) don't have. If you think I think GANs produce higher quality images, I assure you that this is from your imagination as I've never stated such and it is not an opinion that I have. Sorry, I said I wouldn't say this again, so last time for real. [0] There's too much sketchy shit going on in ML research right now and honestly, that's why I've been more passionate about trying to get people to think harder and about context. Again, this is about context. (I also really hate this railroading as it stifles community innovation and sweeps important problems under the rug by saying to just rely on large companies for checkpoints. The community doing so much around Stable Diffusion is awesome, but I want to see that around lots of works because there's a lot that can be accelerated by even a hundredth of this community effort)