The authors derive that under rectified-flow interpolation, the error of the one-step clean prediction equals the realized velocity residual scaled by the noise level.
Self-correcting flow matching via lightweight refiner that reuses hidden states.
CorrFlow achieves consistent improvements in image and video generation, with theoretical guarantees on training efficiency.
Top University
Hangwei Zhang · Nathan Rouillé · Chuanxia Zheng
Nanyang Technological University · Télécom Paris, Institut Polytechnique de Paris
Research Digest··2 min read
The authors introduce CorrFlow, a self-correcting flow-matching framework that uses a lightweight refiner to predict velocity residuals.
§