InvFusion: Bridging Supervised and Zero-shot Diffusion for Inverse Problems

Noam Elata, Hyungjin Chung *, Jong Chul Ye, Tomer Michaeli, Michael Elad · 2025

Read InvFusion: Bridging Supervised and Zero-shot Diffusion for Inverse Problems by Noam Elata, Hyungjin Chung *, Jong Chul Ye, Tomer Michaeli, Michael Elad (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-2504-01689.txt.

Description

Noam Elata and co-authors wrote a 2025 paper on a new sampler for inverse problems, tasks where a clean image must be rebuilt from a blurred or partial one. Diffusion models give strong results for such tasks. Yet past methods split into two camps with a trade-off.

Zero-shot methods take any known blur but lean on shortcuts that cut accuracy. Trained methods model the answer spread in the right way but cannot adapt when the blur changes at test time. Their method is the first trained sampler that takes the blur as an input to the denoiser itself. Tests on two face and object data sets beat rival samplers. The same design also works as a plain error based predictor.

Catalog record

CREATORNoam Elata, Hyungjin Chung *, Jong Chul Ye, Tomer Michaeli, Michael Elad
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE57 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-01689

Browse the full library · About the Tower of Babel