Delta Velocity Rectified Flow for Text-to-Image Editing

Gaspard Beaudouin, Minghan Li, Jaeyeon Kim, Sung-Hoon Yoon, Mengyu Wang · 2025

Read Delta Velocity Rectified Flow for Text-to-Image Editing by Gaspard Beaudouin, Minghan Li, Jaeyeon Kim, Sung-Hoon Yoon, Mengyu Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-05342.txt.

Description

Editing a generated image without oversmoothing it is the target of Gaspard Beaudouin and four co-authors in a 2025 paper on rectified flow models. These generators move noise toward pictures along learned paths. Earlier editing shortcuts often blur fine detail.

Their method, called DRFS, directly models the gap between the source and target velocity fields that steer generation. A shift term that changes over time pushes noisy states closer to the target path. The authors show older editing methods appear as special cases when parts of their method are switched off. On a widely used editing benchmark, DRFS gave better quality, faithfulness, and control with no change to the model itself.

Catalog record

CREATORGaspard Beaudouin, Minghan Li, Jaeyeon Kim, Sung-Hoon Yoon, Mengyu Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-05342

Browse the full library · About the Tower of Babel