Reproduce “SenseFlow: Scaling Distribution Matching for Flow-based Text-to-Image Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xingtong Ge, Xin Zhang, Tongda Xu, Yi Zhang, Xinjie Zhang, Yan Wang, Jun Zhang
SenseFlow introduces a method for scaling Distribution Matching Distillation to large-scale, flow-based text-to-image models, addressing challenges in training stability and timestep importance. The approach enables models like SD 3.5 Large and FLUX.1 dev to achieve superior human-preference scores and strong compositional generalization in a 4-step generation process, often outperforming their multi-step teachers.
Key insights
- Softly aligning the fake distribution model with the generator through proximal updates significantly enhances the stability and scalability of DMD for very large generative models.
- Explicitly capturing and transferring the non-uniform denoising importance within coarse timestep segments from the teacher can substantially improve the student generator's ability to produce high-quality images in few steps.
- Integrating features from pre-trained Vision Foundation Models into the discriminator provides richer, semantically informed adversarial guidance, leading to generated images that align more closely with human preferences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.