The explanation in the original paper turns out not to be true; you can get rid of most of their assumptions and it still works: https://arxiv.org/abs/2208.09392
I’ll admit it is amusing that some assumptions on why it works were incorrect. The core idea of a Markov chain[0] where each state change leads to higher likelihood, is bound to work, even if the rest doesn’t.
In my mind, the Muse paper[1] gets closer to why it works: ultimately, the denoiser tries to match the latent space for an implicit encoder. The Muse system does this more directly and more effectively, by using cross-entropy loss on latent tokens instead.
In a way, the whole problem is no different from a language translation task. The only difference is that the output needs to be decoded into pixels instead of BPE tokens.
I have found in practice that they do not deliver on this front. The loss curve you get is often just a big thick noisy straight line, completely devoid of information about whether it's converging. And the convergence seems to be greatly dependent on model choices and the beta schedule you choose. It's not clear to me at all how to choose those things in a principled manner. Until you train for a long time, you just basically get noise, so it's hard to know when to restart an experiment or keep going. Do I need 10 steps, 100, 1000? I found that training even longer, and longer, and longer, it does get better and better, very slowly, even though this is not displayed in the loss curve, and there seems to be no indication of when the model has "converged" in any meaningful sense. My understanding of why this is the case is that due to the integrative nature of the sampling process, even tiny errors in approximating the noise add up to large divergences.
I've also tried making it conditional on vector quantization codes, and it seems to fail to use them nearly as well as VQGAN does. At least I haven't had much success doing it directly in the diffusion model. After reading more into it, I found that most diffusion-based models actually use a conditional GAN to develop a latent space and a decoder, and the diffusion model is used to generate samples in the latent space. This strikes me that the diffusion model then can never actually do better than the associated GAN's decoder, which surprised me to realize since it's usually proposed as an alternative to GAN.
So, overall I'm failing to grasp the advantages this approach really has over just using a GAN. Obviously it works fantastically for these large scale generative projects, but I don't understand why it's better, to be honest, despite having read every article out there telling me again and again the same things about how it works. E.g. DALLE-1 used VQGAN, not diffusion, and people were pretty wowed by it. I'm not sure why DALLE-2's improvements can be attributed to their change to a diffusion process, if they are still using a GAN to decode the output.
Looking for some intuition, if anyone can offer some. I understand that the nature of how it iteratively improves the image allows it to deduce large-scale and small-scale features progressively, but it seems to me that the many upscaling layers of a large GAN can do the same thing.
My first assumption was that the mdoel I was training was too small: 13 million parameters as opposed to the 1.3 billion in ruDALL-E (not sure how much of this is only the diffusion model). So that's a 100x smaller. I want to experiment with upscaling it.
Reading this I'm wondering if there's more I need to do. For example, training a conditioned model - "cheat" by given it in the index of the Pokemon during training but then you sample without an index - or make the model predict the standard deviation (beta tilde). Or as you say, work with loss functions.
More work to be done here.
> Dalle 1 works thx to the autoregressive model (also no GAN)
It uses an autoregressive model to predict codes for a pretrained VQGAN, doesn't it?
Doesn't Stable Diffusion's autoencoder also use an adversarial loss? Otherwise wouldn't it suffer the typical blurring problems well known to MSE?
I'm just expressing here that my expectation was that this method would be less finicky than GAN because it uses an MSE loss, but unfortunately it seems to have its own difficulties. No silver bullet, I guess. The integration sampling can be quite sensitive to imperfections and diverge easily, at least in early stages of training.
I decided to write this because it feels like the early days of GAN where overall there seems to be lots of these "explain diffusion from scratch" type articles out there, but not yet a lot discussing common pitfalls and how to deal with them.
