Chenyang Yuan

Why do diffusion models produce new images instead of just memorizing the dataset? We show that they learn pixel correlation patterns from the data and therefore denoise locally, which promotes generalization. To test this idea, we compare trained diffusion models with a training-free algorithm that mixes local patches from the dataset. Surprisingly, this simple procedure already reproduces many properties of the trained models. 🧵 Check out this thread for more details about our Spotlight NeurIPS paper with @yuancy, @JustinMSolomon and @vincesitzmann.

At #NeurIPS today in San Diego? Come check out poster #4409 (4:30–7:30 PM) today. We’re excited to share our spotlight paper on the generalization properties of diffusion models. Looking forward to great research conversations! @yuancy @JustinMSolomon @vincesitzmann

I played a small part in the production of this video, and I'm really happy with how it addressed common misconceptions about diffusion models, as well as the beautiful visualizations and animations!

@alec_helbling Thanks, I had a lot of fun building it!

This smalldiffusion project is really fun. It is a tiny library for training and sampling from diffusion models. I love efforts like this, they are much easier to play around with than big libraries like diffusers. Link: github.com/yuanchenyang/s…

In the problem sets, we use the library introduced in the first lecture (github.com/yuanchenyang/s…) to train diffusion models on custom data, as well as using pretrained models as building blocks for a variety of downstream tasks (see examples above) (4/4)

Using score distillation for 3D shape generation (L5) and wrapping up with a summary of the latest research making diffusion models better and faster (L6) The lecture recordings can be found here: youtube.com/playlist?list=… (3/4)

Last month I cotaught a class on diffusion models at MIT during the IAP term: practical-diffusion.org In the lectures, we first introduced diffusion models from a practitioner's perspective, showing how to build a simple but powerful implementation from the ground up (L1) (1/4)