General and Efficient Steering of Unconditional Diffusion

Qingsong Wang, Mikhail Belkin, Yusu Wang, [email protected], [email protected] · 2026

Read General and Efficient Steering of Unconditional Diffusion by Qingsong Wang, Mikhail Belkin, Yusu Wang, [email protected], [email protected] (2026) 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-2602-11395.txt.

Description

Steering diffusion models to unseen terms usually needs retraining or costly gradients. Qingsong Wang and two co-authors present Noise-Aligned RFM Steering, a recipe that steers without gradient guidance at inference. Two offline signals combine: noise matching from principal component statistics of target examples, and trait-machine signal steering from labelled forward signals.

Noise matching gives coarse control at high noise, while the learned direction is reused across middle and late steps through light signal edits. Tests on image benchmarks show better target scores than gradient-based post-hoc guidance, better image quality scores on class-guidance tasks, and large speedups. Code is released for controllable making without retraining.

Catalog record

CREATORQingsong Wang, Mikhail Belkin, Yusu Wang, [email protected], [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-11395

Browse the full library · About the Tower of Babel