Bridging Diffusion Guidance and Anderson Acceleration via Hopfield Dynamics
Kwanyoung Kim · 2026
Read Bridging Diffusion Guidance and Anderson Acceleration via Hopfield Dynamics by Kwanyoung Kim (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-2603-02531.txt.
Description
Kwanyoung Kim links diffusion guidance to a classic acceleration trick in this 2026 paper. Guidance improves diffusion image quality by pushing apart two model outputs. It costs extra passes at run time and fails for distilled, single-step models. Attention-space methods are cheaper, but their theory was murky.
Kim models attention dynamics as fixed-point steps inside modern Hopfield networks. Attention extrapolation then becomes a special case of Anderson acceleration on those dynamics. A weak contraction property of the dynamics supports that reading. His method splits attention updates into parts along and across the guidance direction. That split stabilizes the speedup and squeezes the most from guidance. The plug-and-play method fits existing frameworks and improves generation quality a lot.
Catalog record
| CREATOR | Kwanyoung Kim |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66588 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.02531