4 ms·

>Stop buying into hype. There isn't one model to rule them all, there are models that are better in differing contexts. I didn't say any of that, there is simp

by GaggiX 3y ago

>Stop buying into hype. There isn't one model to rule them all, there are models that are better in differing contexts. I didn't say any of that, there is simply no open source GAN model that can compete with the open source diffusion models we have today, and the fact that these models can be distilled down to 1/2/4 steps makes GANs less attractive.

godelski 3y ago
> I didn't say any of that I'm sorry, then I don't know what you're saying. Because what I read was 2 claims. 1) GAN quality is less than diffusion. 2) GANs can't do T2I. I think I adequately showed that both these assumptions were wrong. I'm not sure what else your comment meant as that contains all of its words... > there is simply no open source GAN model that can compete with the open source diffusion models we have today Again, I provided citations. Do you want the github links? Here's GigaGAN's: https://github.com/mingukkang/GigaGAN https://github.com/mingukkang/GigaGAN (checkpoints under the evaluation folder). What do you mean by compete? Because the metric scores are quite comparable. I think that's a reasonable interpretation of the word "competitive" but I'm a generative researcher so we might be using the terms to mean different things. (I even like diffusion more fwiw, but I'm particularly more interested in tractable density models) Do you mean "the open community has rallied around Stable Diffusion and sunk in more time to tuning this model and producing textual inversions and LoRAs which far out surpass that of any other model, even including other diffusion models and is a phenomena nearly exclusive to Stable Diffusion and has been mostly accomplished in the last year"? Because if so, yeah, I agree. But that's not a really good argument for saying Diffusion is better than GANs. It's a completely different argument. That's just to say __Stable Diffusion__ (not the class of diffusion models, or even more specifically latent diffusion models (the two are different)) has better community support. That's a reasonable argument, but a different one. Because it really just shows what you can do to __any__ model. These techniques and efforts are not architecture dependent nor are they even mode dependent. We see similar community effort around OpenLLaMa and GPT but less so around Bard/Gemini or Claude. So what? That's not really relevant to the conversation nor to the specific context that we're discussing which is very likely not going to include many of those TIs or LoRAs (or if it would you'd probably be training a custom one, so the point is again moot since LoRAs are not unique to diffusion models). I'm really happy a lot of people have entered the community and are effectively doing research, but it'd be quite naive to say that such a thing isn't possible around other architectures. Evangelism is quite useful, but not if it turns into religious beliefs. Hell, we can argue that Linux is better than Windows because Linux has a bigger hacker community but I'm not sure that's a great or even meaningful argument because it's void of context in what better actually means. The better OS is clearly situationally dependent and makes a lot of the OS holy wars silly. > and the fact that these models can be distilled Literally any model can be distilled. I'm not understanding your argument. Are you just arguing that such efforts have already taken place? Sure, I'll agree to that. But it's worth noting that the distilled diffusion models are still quite large and much larger than some of the aforementioned works. > 1/2/4 steps makes GANs less attractive. I even addressed the one step process. Which, single step diffusion generation represents a significant decrease in quality, so this seems to run counter to your prior argument. This is why I'm a bit confused. If we're going with a low step diffusion model then the case for the GAN becomes clearer because even your non-distilled GAN's inference is still higher and it's quality is definitely superior. So really, I am confused. I'm not sure what you're arguing. We can get into the weeds and discuss diversity, recall, memorization, density estimation, and all manner of things but these are quite open questions and frankly understudied. Plus we'll have to be extra nuanced because the metrics are proxies and incorporate different biases that different architectures are going to suffer from, making it difficult to compare in a more fair sense. But that's okay. Let's also be clear: GAN != StyleGAN{1,2,2-ada,3,XL,T,*} and diffusion != StableDiffusion. There are other GANs and other diffusions and even other image synthesis models, the vast majority of them being open source. Again, I'll assert, there are no universal models that are best for all situations. There are only models that are best at specific situations. You can translate this to "AGI doesn't exist but we have narrow AI" if that is clearer. But my claim is a bit broader still because if that AGI took a warehouse to run it still wouldn't be contextually appropriate.
- ShamelessC 3y ago
  Just FYI, it will less obvious that you’re entrenched and “dug in” to the field of GAN research if you’re less defensive when people say fairly reasonable things about different ideas.
  - godelski 3y ago
    Problem is none of what you believe about me is true. By main focus area is explicit density estimation, GANs are implicit. I like math... But you can check my comment history to see that I really hate hype. I said there are better models for different tasks. Here's a breakdown: - Autoregressive: Best for time sequential data - Normalizing Flows: Best for density estimation and statistics - Diffusion: Best at general image synthesis, editing, and diversity. - GANs: Best when working on edge devices or throughput is critical - VAEs: Best for situations between diffusion and flows, where a implicit PCA is desired You can tell why I discussed GANs because it hit the specific use case. Remember my first comment is literally saying the motivation is trading image quality and diversity for generation time. It's because again, the claim is that there isn't one size fits all models. Such a notion is silly. You don't use a diffusion model to do real time image upscaling (e.g. for video) nor do you want it for RTX supersampling, but you do want diffusion for general image synthesis tasks in such areas like photoshop. Inpainting, outpainting, T2I, unconditional, I2I, and all that you want diffusion because it is much better for those tasks. Just... diffusion requires big models and they're slow and the raspberry pi is small. So you make sacrifices.
- diimdeep 3y ago
  Sorry, but at the link you provided there is no model weights, only inferred images offered for examination, so at this point it is at vaporware stage compared to stable diffusion tools.
  - godelski 3y ago
    Sorry, you're right. That is a really weird thing to do. Just provide 6GB of images...? They even don't have an issues page. Looks to be a common thing by that author. You're right, very suspicious. Lafite has their checkpoints at least. Results aren't great, but it is small and fast. They have a colab but it's broken (lol). Fix by removing the torch versions, and add gdown. Replace the wget line with "!gdown https://drive.google.com/u/0/uc?id=17ER7Yl02Y6yCPbyWxK_tGrJ8RKkcieKq https://drive.google.com/u/0/uc?id=17ER7Yl02Y6yCPbyWxK_tGrJ8..." (checkpoint from from their github). Then everything will work fine. It took some time to get some decent outputs (but then again so did my first time with any diffusion model. This is definitely lower quality though). At least the authors look engaged in the github issues and do show how to get better results. (Always be suspicious of the images shown in papers... this one is certainly no exception) https://github.com/drboog/Lafite https://github.com/drboog/Lafite

by GaggiX 3y ago

godelski 3y ago

> I didn't say any of that I'm sorry, then I don't know what you're saying. Because what I read was 2 claims. 1) GAN quality is less than diffusion. 2) GANs can't do T2I. I think I adequately showed that both these assumptions were wrong. I'm not sure what else your comment meant as that contains all of its words... > there is simply no open source GAN model that can compete with the open source diffusion models we have today Again, I provided citations. Do you want the github links? Here's GigaGAN's: https://github.com/mingukkang/GigaGAN https://github.com/mingukkang/GigaGAN (checkpoints under the evaluation folder). What do you mean by compete? Because the metric scores are quite comparable. I think that's a reasonable interpretation of the word "competitive" but I'm a generative researcher so we might be using the terms to mean different things. (I even like diffusion more fwiw, but I'm particularly more interested in tractable density models) Do you mean "the open community has rallied around Stable Diffusion and sunk in more time to tuning this model and producing textual inversions and LoRAs which far out surpass that of any other model, even including other diffusion models and is a phenomena nearly exclusive to Stable Diffusion and has been mostly accomplished in the last year"? Because if so, yeah, I agree. But that's not a really good argument for saying Diffusion is better than GANs. It's a completely different argument. That's just to say __Stable Diffusion__ (not the class of diffusion models, or even more specifically latent diffusion models (the two are different)) has better community support. That's a reasonable argument, but a different one. Because it really just shows what you can do to __any__ model. These techniques and efforts are not architecture dependent nor are they even mode dependent. We see similar community effort around OpenLLaMa and GPT but less so around Bard/Gemini or Claude. So what? That's not really relevant to the conversation nor to the specific context that we're discussing which is very likely not going to include many of those TIs or LoRAs (or if it would you'd probably be training a custom one, so the point is again moot since LoRAs are not unique to diffusion models). I'm really happy a lot of people have entered the community and are effectively doing research, but it'd be quite naive to say that such a thing isn't possible around other architectures. Evangelism is quite useful, but not if it turns into religious beliefs. Hell, we can argue that Linux is better than Windows because Linux has a bigger hacker community but I'm not sure that's a great or even meaningful argument because it's void of context in what better actually means. The better OS is clearly situationally dependent and makes a lot of the OS holy wars silly. > and the fact that these models can be distilled Literally any model can be distilled. I'm not understanding your argument. Are you just arguing that such efforts have already taken place? Sure, I'll agree to that. But it's worth noting that the distilled diffusion models are still quite large and much larger than some of the aforementioned works. > 1/2/4 steps makes GANs less attractive. I even addressed the one step process. Which, single step diffusion generation represents a significant decrease in quality, so this seems to run counter to your prior argument. This is why I'm a bit confused. If we're going with a low step diffusion model then the case for the GAN becomes clearer because even your non-distilled GAN's inference is still higher and it's quality is definitely superior. So really, I am confused. I'm not sure what you're arguing. We can get into the weeds and discuss diversity, recall, memorization, density estimation, and all manner of things but these are quite open questions and frankly understudied. Plus we'll have to be extra nuanced because the metrics are proxies and incorporate different biases that different architectures are going to suffer from, making it difficult to compare in a more fair sense. But that's okay. Let's also be clear: GAN != StyleGAN{1,2,2-ada,3,XL,T,*} and diffusion != StableDiffusion. There are other GANs and other diffusions and even other image synthesis models, the vast majority of them being open source. Again, I'll assert, there are no universal models that are best for all situations. There are only models that are best at specific situations. You can translate this to "AGI doesn't exist but we have narrow AI" if that is clearer. But my claim is a bit broader still because if that AGI took a warehouse to run it still wouldn't be contextually appropriate.

ShamelessC 3y ago
Just FYI, it will less obvious that you’re entrenched and “dug in” to the field of GAN research if you’re less defensive when people say fairly reasonable things about different ideas.
- godelski 3y ago
  Problem is none of what you believe about me is true. By main focus area is explicit density estimation, GANs are implicit. I like math... But you can check my comment history to see that I really hate hype. I said there are better models for different tasks. Here's a breakdown: - Autoregressive: Best for time sequential data - Normalizing Flows: Best for density estimation and statistics - Diffusion: Best at general image synthesis, editing, and diversity. - GANs: Best when working on edge devices or throughput is critical - VAEs: Best for situations between diffusion and flows, where a implicit PCA is desired You can tell why I discussed GANs because it hit the specific use case. Remember my first comment is literally saying the motivation is trading image quality and diversity for generation time. It's because again, the claim is that there isn't one size fits all models. Such a notion is silly. You don't use a diffusion model to do real time image upscaling (e.g. for video) nor do you want it for RTX supersampling, but you do want diffusion for general image synthesis tasks in such areas like photoshop. Inpainting, outpainting, T2I, unconditional, I2I, and all that you want diffusion because it is much better for those tasks. Just... diffusion requires big models and they're slow and the raspberry pi is small. So you make sacrifices.
diimdeep 3y ago
Sorry, but at the link you provided there is no model weights, only inferred images offered for examination, so at this point it is at vaporware stage compared to stable diffusion tools.
- godelski 3y ago
  Sorry, you're right. That is a really weird thing to do. Just provide 6GB of images...? They even don't have an issues page. Looks to be a common thing by that author. You're right, very suspicious. Lafite has their checkpoints at least. Results aren't great, but it is small and fast. They have a colab but it's broken (lol). Fix by removing the torch versions, and add gdown. Replace the wget line with "!gdown https://drive.google.com/u/0/uc?id=17ER7Yl02Y6yCPbyWxK_tGrJ8RKkcieKq https://drive.google.com/u/0/uc?id=17ER7Yl02Y6yCPbyWxK_tGrJ8..." (checkpoint from from their github). Then everything will work fine. It took some time to get some decent outputs (but then again so did my first time with any diffusion model. This is definitely lower quality though). At least the authors look engaged in the github issues and do show how to get better results. (Always be suspicious of the images shown in papers... this one is certainly no exception) https://github.com/drboog/Lafite https://github.com/drboog/Lafite