We explore a new class of diffusion models based on the transformer architecture. We train latent diffusion models of images, replacing the commonly-used U-Net backbone with a transformer that operates on latent patches. We analyze the scalability of our Diffusion Transformers (DiTs) through the lens of forward pass complexity as measured by Gflops. We find that DiTs with higher Gflops -- through increased transformer depth/width or increased number of input tokens -- consistently have lower FID. In addition to possessing good scalability properties, our largest DiT-XL/2 models outperform all prior diffusion models on the class-conditional ImageNet 512x512 and 256x256 benchmarks, achieving a state-of-the-art FID of 2.27 on the latter.
Metadata: arXiv. Abstracts retain their original authors’ rights.
In the conversation
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
It's a bit more complicated than that and I don't think you're being fair.
StarFlow and the AR models are fixed but DiT is being compared at different amount of steps and we don't really care if we generate garbage at blazing speeds[0]. Go look at... also Figure 10 (lol) from the DiT paper[1], it compares FID to model sizes and sampling steps. It looks like StarFlow is comparing to DiT-XL/2-G. In [1] they do {16,32,64,128,256,1024} steps which corresponds to (roughly) 10k-FID of 60, 35, 25, 22, 21, 20. Translating to StarFlow's graph we'll guesstimate 21,23,50. There's a big difference between 50 and 23 but what might surprise you is that there's a big difference between…
Read the full comment
25 and 20. Remember that this is a metric that is lower bounded, and that lower bound is not 0... You also start running into the limitations of the metric the closer you get to its lower bound, adding another layer of complexity when comparing[2]
The images from the paper (I believe) are all at 250 steps, which StarFlow is beating at a batch size of 4. So let's look at batches and invert the data. It is imgs/sec so let's do (1/<guestimate of y-value>) * batch. We get this
Batch DiT SF 1 10s 20s 2 20s 30s 4 40s 30s 8 80s 30s 16 160s 30s ... So what's happening here is that StarFlow is invariant to the batch size while DiT is not. Obviously this won't hold forever, but DiT doesn't get advantage from batching. You could probably make up these differences by caching the model because it looks like there's a turn from model loading dominating to actual generation dominating. Whereas StarFlow has that turnover at batch 2.
And batching (even small batches) is going to be pretty common, especially when talking about industry. The scaling here is a huge win to them. It (roughly) costs you just as much to generate 64 images as it does 2. Worst case scenario, you hand your customers batched outputs and they end up happier because frankly, generating images is still an iterative process and good luck getting the thing you want on just the first shot even if you got all your parameters dialed in. So yeah, that makes a much better product.
I'll also add 2 things. 1) you can get WAY more compression out of Normalizing Flows 2) there's just a ton you can do with Flows that you can't with diffusion. The explicit density isn't helpful only for the math nerds. It is helpful for editing, concept segmentation, interpolation, interpretation, and so much more.
[0] https://tvgag.com/content/quotes/6004-jpg.jpg
[1] https://arxiv.org/abs/2212.09748
[2] Basically, place exponentially growing importance on FID gaps as they lower and then abandon the importance completely because it doesn't matter. As an example, take FFHQ-256 with FID-50k. Image quality difference between 50 and 20 is really not that big, visually. But there's a *HUGE** difference between 10 and 5. Visually, probably just as big as the difference between 5 and 3. But once you start going below 3 you really shouldn't rely on the metric anymore and comparing a 2.5 model to 2.7 is difficult.
Image diffusion models also use transformers nowadays. Here is the original "diffusion transformer" paper: https://arxiv.org/abs/2212.09748
Earlier image diffusion models used U-nets: https://en.wikipedia.org/wiki/U-Net
I understand that Sora is very popular, so it makes sense to refer to it, but when saying it is similar to Sora, I guess it actually makes more sense to say that it uses a Diffusion Transformer (DiT) (https://arxiv.org/abs/2212.09748) like Sora. We don't really know more details on Sora, while the original DiT has all the details.
On the other hand, this is the first convincing use of a “diffusion transformer” [1]. My understanding is that videos and images are tokenized into patches, through a process that compresses the video/images into abstracted concepts in latent space. Those patches (image/video concepts in latent space) can then be used with transformers (because patches are the tokens). The point is that there is plenty of room for optimization following the first demonstration of a new architecture.
Edit: sorry, it’s not the first diffusion transformer. That would be [2]
[1] https://openai.com/research/video-generation-models-as-world...
[2] https://arxiv.org/abs/2212.09748
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference between training and test performance.
In honor of Alan Turing's hundredth birthday, I unwisely set out some thoughts about one of Turing's obsessions throughout his life, the question of physics and free will.