Few-step diffusion model field is wild, and there are many methods trying to train a high-quality few-step generator from scratch: Consistency Models, Shortcut Models, and MeanFlow. Turns out, they could be unified in a quite elegant way, which we did in our recent work.

The work was done during my internship @Snap. Thanks to my amazing collaborators @isskoro, @siarohin9013, @WilliMenapace, Michael Vasilkovsky, @qu_1006, @SergeyTulyakov

We proposed AlphaFlow, a new framework that unifies prior training losses and uses a curriculum strategy to achieve better performance. AlphaFlow-XL/2+ hits new SOTA on ImageNet-1K under vanilla DiT backbone. Paper: https://t.co/GGsldkk91K Code:

@NeurIPSConf Just arrived in Vancouver for #NeurIPS2024! (Lost my luggages AGAIN😂) I will be presenting several works, focusing on #DiffusionModels and #RepresentationLearning. Looking forward to speaking to everyone! See lists below:

Our work explores the locally linear and low-rank feature spaces in diffusion models, which further allows localized, linear, transferable, and composable image editing. It is applicable to various models and can be utilized both in unsupervised and text-supervised way.

Happy to share that our work on interpretable and localized image editing on diffusion model is accepted by #neurips2024! Many thanks to my teammates and advisor. https://t.co/8oGDdeaxCN