CorrFlow: Self-Correcting Flow Matching

Key points

• Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them.1
• Under rectified-flow interpolation, the error of the one-step clean prediction equals the realized velocity residual scaled by the noise level.1
• Our theoretical analyses establish conditions under which feature reuse outperforms raw-input affine predictors and refiner training achieves greater residual-risk reduction than continued base-model training under a matched training compute budget.1
• Experiments on class-conditional image generation, text-to-video generation and image-to-video generation show consistent improvements over the corresponding base models, including gains in visual quality and physical consistency.1

Sources (1)

• [1]CorrFlow: Self-Correcting Flow Matching
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 03:35 PM

  We introduce CorrFlow, a self-correcting flow-matching framework that reuses this information through a lightweight refiner.

  Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them.

Extractive summary: sentences quoted from the sources.