SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers

Javad Rajabi, Kimia Shaban, Koorosh Roohi, David B. Lindell, Babak Taati, Vector Institute · 2026

Read SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers by Javad Rajabi, Kimia Shaban, Koorosh Roohi, David B. Lindell, Babak Taati, Vector Institute (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-2605-22668.txt.

Description

Rajabi and others tackle the drop in diffusion transformer image quality when generating above training resolution. Older training free fixes scale rotary position embeddings uniformly, trading global structure against fine detail. Their SEGA method instead scales attention across rotary parts adaptively, guided by the latent image's spatial frequency pattern at each denoising step. This scaling keeps both overall coherence and fine detail. Tests show SEGA beats top training free baselines across several target resolutions.

Catalog record

CREATORJavad Rajabi, Kimia Shaban, Koorosh Roohi, David B. Lindell, Babak Taati, Vector Institute
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67402
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.22668

Browse the full library · About the Tower of Babel