LUSD: Localized Update Score Distillation for Text-Guided Image Editing

Worameth Chinchuthakun 1 Tossaporn Saengja 1,2 Nontawat Tritrong 1 Pitchaporn Rewatbowornwong 1 Pramook Khungurn 3 Supasorn Suwajanakorn 1 1 VISTEC 2 SiData+ 3 Pixiv · 2025

Read LUSD: Localized Update Score Distillation for Text-Guided Image Editing by Worameth Chinchuthakun 1 Tossaporn Saengja 1,2 Nontawat Tritrong 1 Pitchaporn Rewatbowornwong 1 Pramook Khungurn 3 Supasorn Suwajanakorn 1 1 VISTEC 2 SiData+ 3 Pixiv (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-2503-11054.txt.

Description

Worameth Chinchuthakun and co-authors introduce a 2025 method for editing images with text prompts while leaving the rest of the picture alone. Their tool, called LUSD, builds on score distillation, a way of borrowing the skill of a text-to-image model without new training, but older tools often fail at jobs like adding an object.

Looking into those failures, the team found that update steps vary wildly in size and place across the image. They add two fixes: a spatial rule based on attention maps, and a step that filters and scales the updates. In tests, edited images matched the prompt more often, users picked their results over rival tools, and the untouched parts of each scene stayed intact.

Catalog record

CREATORWorameth Chinchuthakun 1 Tossaporn Saengja 1,2 Nontawat Tritrong 1 Pitchaporn Rewatbowornwong 1 Pramook Khungurn 3 Supasorn Suwajanakorn 1 1 VISTEC 2 SiData+ 3 Pixiv
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE66 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-11054

Browse the full library · About the Tower of Babel