Hair in a stream of air
Prompt A person uses a low heat setting on the hairdryer to gently dry their fine hair.
Compare CorrFlow with
CorrFlowSelf-Correcting Flow Matching
1 PVG, Nanyang Technological University2 Télécom Paris, Institut Polytechnique de Paris
THE IDEA
Abstract
CorrFlow learns to correct its own predictions. A lightweight refiner reuses the base model’s clean prediction, hidden states, and conditioning to refine the velocity before each sampling update—with no additional base-model evaluations.
Read the full abstract
Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them. We introduce CorrFlow, a self-correcting flow-matching framework that reuses this information through a lightweight refiner. Under rectified-flow interpolation, the error of the one-step clean prediction equals the realized velocity residual scaled by the noise level. Motivated by this identity, the refiner learns to predict residual corrections from the base model’s clean prediction, conditioning, noise level and hidden states from the same forward pass. The correction is added to the base velocity before each sampling update, so subsequent predictions are made from the corrected state. This requires no additional base-model evaluations and only one lightweight refiner evaluation per step. Our theory gives conditions under which feature reuse outperforms raw-input affine predictors, refiner training outperforms continued base-model training at matched compute, and deployed correction lowers velocity error under its own sampler law. 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.
01 / VIDEO GENERATION
See the difference in time.
Explore physical interactions with VideoPhy2 and Physics-IQ,
followed by FinePhyEval and VBench.
Chapter 01 / Text to video
VideoPhy2
Everyday actions and physical interactions.
Choose a video case
Select a thumbnail to explore a different scene.
A somersault on a bouncy castle
Prompt A child somersaults on a large bouncy castle, their body rotating fully in the air.
Compare CorrFlow with
Paddling across still water
Prompt A canoeist uses a single-bladed paddle to propel their canoe across a lake, the paddle's movement visible against the still water.
Compare CorrFlow with
Jet skis and colliding wakes
Prompt Two jetskis race side-by-side, their wakes colliding in a visible spray of water.
Compare CorrFlow with
Parasailing over the ocean
Prompt A parasailer, harnessed to a bright orange parachute, is pulled along a calm ocean surface by a speedboat.
Compare CorrFlow with
A coin tossed into a cup
Prompt A coin is flung into the air and caught in a metal cup.
Compare CorrFlow with
Chapter 02 / Image to video
Physics-IQ
Image-conditioned physical events.
Choose a video case
Select a thumbnail to explore a different scene.
Ball ramp
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. A simple ramp made of cardboard, propped up by a blue block on a light-colored wooden table. There's a black pipe to the left of the frame that directly points towards it. The yellow tennis ball rolls straight out of the black pipe onto the ramp.
Compare CorrFlow with
Ball rolls on glass
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. A piece of clear glass resting on the edge of a light-colored wooden table against a plain white wall. The blue tennis ball rolls from the wooden table over the glass.
Compare CorrFlow with
Cut orange
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. One half of a tangerine has been placed on a glass cutting board. The knife slices through the tangerine.
Compare CorrFlow with
Domino in juice
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. The mug is sitting on a wooden surface. The blue-black grabber tool drops the white ceramic domino into the mug that is filled with a dark-colored liquid.
Compare CorrFlow with
Light on statue
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. The small porcelain statue sits on the black, clockwise-rotating platform, and the spotlight points toward the porcelain statue. The small porcelain statue rotates clockwise on the black platform while the spotlight casts a large shadow on the wall behind it.
Compare CorrFlow with
Match
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. The lit match is being lowered into the glass of water.
Compare CorrFlow with
Mirror ball fall
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. The tennis ball attached to a magnet that hangs from an orange cord was hovering in the air in front of the mirror. The mirror creates the illusion of two tennis balls. The tennis ball hanging in front of the mirror is slowly lowered.
Compare CorrFlow with
Paper fall water
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. The orange-black grabber tool is suspended over the bowl of water on a wooden table. The orange-black grabber tool releases the crumpled piece of paper over the bowl of water.
Compare CorrFlow with
Paper smoke
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. The piece of folded paper sits on a glass cutting board. The piece of folded paper is burning while white smoke is emitting from it.
Compare CorrFlow with
Rolling reflection
Prompt Static locked-off single-shot with fixed frame throughout filmed with constant framerate in real-time. The scene shows a realistic scientific demonstration. The scene only contains the described setup and actions. The 30lb kettlebell sits next to a mirror that reflects the kettlebell on a wooden table. The tennis ball rolls slowly towards the 30lb kettlebell next to a mirror and hits the kettlebell.
Compare CorrFlow with
Chapter 03 / Text to video
FinePhyEval
Object interactions and physical events.
Choose a video case
Select a thumbnail to explore a different scene.
A domino dropped into liquid
Prompt A grabber tool holding a white domino drops the domino into a dark-colored liquid in a blue mug that is on a wooden surface. Static shot with no camera movement.
Compare CorrFlow with
Dominoes and a rubber duck
Prompt A yellow rubber duck is positioned in the middle of a line of black and white dominoes on a wooden table. A stick attached to a black rotating platform rotates clockwise and knocks the first domino block. Static shot with no camera movement.
Compare CorrFlow with
Chapter 04 / Text to video
VBench
Appearance, composition, and motion.
Choose a video case
Select a thumbnail to explore a different scene.
A person is bending metal
Prompt A person is bending metal
Compare CorrFlow with
A person is ironing
Prompt A person is ironing
Compare CorrFlow with
A chair and a couch
Prompt a chair and a couch
Compare CorrFlow with
A knife and a tv
Prompt a knife and a tv
Compare CorrFlow with
A skateboard and a surfboard
Prompt a skateboard and a surfboard
Compare CorrFlow with
A backpack
Prompt a backpack
Compare CorrFlow with
A handbag
Prompt a handbag
Compare CorrFlow with
A beautiful coastal beach in spring, waves lapping on sand by Vincent van Gogh
Prompt A beautiful coastal beach in spring, waves lapping on sand by Vincent van Gogh
Compare CorrFlow with
Scene consistency
Prompt A boat sailing leisurely along the Seine River with the Eiffel Tower in background by Vincent van Gogh
Compare CorrFlow with
A panda standing on a surfboard in the ocean in sunset
Prompt A panda standing on a surfboard in the ocean in sunset.
Compare CorrFlow with
Scene consistency
Prompt Snow rocky mountains peaks canyon. snow blanketed rocky mountains surround and shadow deep canyons. the canyons twist and bend through the high elevated mountain peaks.
Compare CorrFlow with
Amusement park
Prompt amusement park
Compare CorrFlow with
A sports ball on the right of a baseball bat, front view
Prompt a sports ball on the right of a baseball bat, front view
Compare CorrFlow with
An orange on the top of a carrot, front view
Prompt an orange on the top of a carrot, front view
Compare CorrFlow with
All videos are from the paper’s supplementary material. “Play together” restarts the selected pair; the shorter clip holds its last frame until both finish. Individual video controls work independently.
02 / IMAGE GENERATION
ImageNet 256 × 256
Class-conditional image generation.
Small details, varied
subjects, rich textures.
BUILD ON THIS WORK
Citation
If you find CorrFlow useful in your research,
please consider
citing our work.
BibTeX
@misc{zhang2026corrflow,
title={CorrFlow: Self-Correcting Flow Matching},
author={Hangwei Zhang and Nathan Rouillé and Chuanxia Zheng},
year={2026},
eprint={2610.24750},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.24750},
}