Faster convergence for diffusion models(sihyun.me)
sihyun.me
Faster convergence for diffusion models
https://sihyun.me/REPA/
4 comments
The title is not wrong, but it also doesn't feel correct either. What they do here is they use a pre-trained model to guide the training of a 2nd model. Of course, that massively speeds up training of the 2nd model. But it's not like you can now train a diffusion model from scratch 20x faster. Instead, this is a technique for transplanting an existing model onto a different architecture so that you don't have to start training from 0.
It does feel right to me, because it's not distilling the second model, and in fact the second model is not an image generation model at all, but a visual encoder. That is, it's a more "general purpose" model which specializes in extracting semantic information from images.
In hindsight it makes total sense - generative image models don't automatically start out with an idea of semantic meaning or the world, and so they have to implicitly learn one during training. That's a hard task by itself, and it's not specifically trained for this task, but rather learns it on the go at the same time as the network learns to create images. The idea of the paper then is to provide the diffusion model with a preexisting concept of the world by nudging its internal representations to be similar to the visual encoders'. As I understand DINO isn't even used during inference after the model is ready, it's just about representations.
I wouldn't at all describe it as "a technique for transplanting an existing model onto a different architecture". It's different from distillation because again, DINO isn't an image generation model at all. It's more like (very roughly simplifying for the sake of analogy) instead of teaching someone to cook from scratch, we're starting with a chef who already knows all about ingredients, flavors, and cooking techniques, but hasn't yet learned to create dishes. This chef would likely learn to create new recipes much faster and more effectively than someone starting from zero knowledge about food. It's different from telling them to just copy another chef's recipes.
In hindsight it makes total sense - generative image models don't automatically start out with an idea of semantic meaning or the world, and so they have to implicitly learn one during training. That's a hard task by itself, and it's not specifically trained for this task, but rather learns it on the go at the same time as the network learns to create images. The idea of the paper then is to provide the diffusion model with a preexisting concept of the world by nudging its internal representations to be similar to the visual encoders'. As I understand DINO isn't even used during inference after the model is ready, it's just about representations.
I wouldn't at all describe it as "a technique for transplanting an existing model onto a different architecture". It's different from distillation because again, DINO isn't an image generation model at all. It's more like (very roughly simplifying for the sake of analogy) instead of teaching someone to cook from scratch, we're starting with a chef who already knows all about ingredients, flavors, and cooking techniques, but hasn't yet learned to create dishes. This chef would likely learn to create new recipes much faster and more effectively than someone starting from zero knowledge about food. It's different from telling them to just copy another chef's recipes.
The technique in this paper would still be rightly described as distillation. In this case it's distillation of "internal" representations rather than the final prediction. This a reasonably common form of distillation. The interesting observation in this paper is that including an auxiliary distillation loss based on features from a non-generative model can be beneficial when training a generative model. This observation leads to interesting questions like, eg, which parts of the overall task of generating images (diffusionly) are being learned faster/better due to this auxiliary distillation loss.
Yes, now it seems obvious, but before this it wasn't clear that that would be something that could speed things up, due to the fact that the pretrained model was trained on a separate objective. It's a brilliant idea that works amazingly.
To me, it seemed that the technique presented here was just a logical continuation of methods that OpenAI used when they trained the Dota agents:
https://arxiv.org/pdf/1912.06719v1
And, arguably, Facebook's unsupervised pre-training for their multi-modal speech-to-text models is kind of the same idea as unsupervised pre-training for a multi-modal text-to-image diffuser.
https://ai.meta.com/research/publications/wav2vec-2.0-a-fram...
https://arxiv.org/pdf/1912.06719v1
And, arguably, Facebook's unsupervised pre-training for their multi-modal speech-to-text models is kind of the same idea as unsupervised pre-training for a multi-modal text-to-image diffuser.
https://ai.meta.com/research/publications/wav2vec-2.0-a-fram...
Diffusion models are already being evaluated using pretrained SSL models à la CLIP Score [1]. So it makes sense that one would incorporate that directly into training the model from scratch.
[1] https://huggingface.co/docs/diffusers/en/conceptual/evaluati...
[1] https://huggingface.co/docs/diffusers/en/conceptual/evaluati...
I wonder how well this technique works if the distribution of the training dataset between the diffusion model and the image encoder is quite different, for example if you use DinoV2 as the encoder but train the diffusion model on anime.
Still waiting for a competitive diffusion llm
So I can't find that paper that was posted on HN that said that, when viewed under the right theoretical framework, asserts that diffusion and transformers are doing the same thing under a different basis.. am I misrembering something?
Why?
Diffusion works significantly better for images than sequential pixel generation, there is a good chance it would work better for language as well.
Sequential generation used to be state of the art in 2016 and it's basically how current LLMs work:
https://arxiv.org/abs/1601.06759
Sequential generation used to be state of the art in 2016 and it's basically how current LLMs work:
https://arxiv.org/abs/1601.06759
Neural LMs used to be based on recurrent architectures until the Transformer came along. That architecture is not recursive.
I am not sure that a diffusion approach is all that suitable for generating language. Word are much more discrete than pixels.
I am not sure that a diffusion approach is all that suitable for generating language. Word are much more discrete than pixels.
I meant sequential generation, I didn't mean using an RNN.
Diffusion doesn't work on pixels directly either, it works on a latent representation.
Diffusion doesn't work on pixels directly either, it works on a latent representation.
The most popular method using autoregression in image generation space is to predict image patches/tokens and not pixels, though that still scales worse than diffusion.
A fairly new but promising approach for autoregression that seems to scale as well as diffusion is predicting the next image scale/resolution rather than the next image patch.
https://arxiv.org/abs/2404.02905
A fairly new but promising approach for autoregression that seems to scale as well as diffusion is predicting the next image scale/resolution rather than the next image patch.
https://arxiv.org/abs/2404.02905
I had similar thoughts to you.
However diffusion models suck at details, like how many fingers on a hand, and with language words and characters matter, both which ones and where they are.
So while I'm sure diffusion could produce walls of text that look convincingly like a blog post at a glance say, I'm not sure it would hold up to anyone actually reading.
However diffusion models suck at details, like how many fingers on a hand, and with language words and characters matter, both which ones and where they are.
So while I'm sure diffusion could produce walls of text that look convincingly like a blog post at a glance say, I'm not sure it would hold up to anyone actually reading.