Reproduce “ProReflow: Progressive Reflow with Decomposed Velocity” 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

Lei Ke, Haohang Xu, Xuefei Ning, Yu Li, Jiajun Li, Haoling Li, Yuxuan Lin, Dongsheng Jiang and 2 more

ProReflow develops a refined training framework for accelerating flow-based generative models, achieving state-of-the-art performance in few-step image generation while significantly lowering training costs. The method utilizes a progressive reflow strategy for trajectory approximation and an aligned v-prediction objective that prioritizes directional accuracy.

Key insights

  • Approximating complex diffusion trajectories is more tractable when broken down into a progressive, multi-stage curriculum learning process that gradually reduces the number of segments.
  • Directional accuracy in velocity prediction is significantly more critical for maintaining image generation quality than magnitude precision in flow-based models.
  • The progressive training approach, supported by knowledge distillation theory, leads to smaller approximation errors and higher effective learning rates.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.