Saved in:
Bibliographic Details
Main Authors: Rajabi, Javad, Shaban, Kimia, Roohi, Koorosh, Lindell, David B., Taati, Babak
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.22668
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910246526517248
author Rajabi, Javad
Shaban, Kimia
Roohi, Koorosh
Lindell, David B.
Taati, Babak
author_facet Rajabi, Javad
Shaban, Kimia
Roohi, Koorosh
Lindell, David B.
Taati, Babak
contents Diffusion transformers (DiTs) have emerged as a dominant architecture for text-to-image generation, yet their performance drops when generating at resolutions beyond their training range. Existing training-free approaches mitigate this by modifying inference-time attention behavior, often through Rotary Position Embeddings (RoPE) extrapolation combined with attention scaling. However, these strategies apply a uniform and content-agnostic scaling across RoPE components with distinct frequency characteristics, inducing a trade-off between preserving global structure and recovering fine detail. We introduce SEGA, a training-free method that dynamically scales attention across RoPE components according to the latent's spatial-frequency structure at each denoising step. This adaptive scaling improves both structural coherence and fine-detail fidelity. Experiments show that SEGA consistently improves high-resolution synthesis across multiple target resolutions, outperforming state-of-the-art training-free baselines.
format Preprint
id arxiv_https___arxiv_org_abs_2605_22668
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers
Rajabi, Javad
Shaban, Kimia
Roohi, Koorosh
Lindell, David B.
Taati, Babak
Computer Vision and Pattern Recognition
Diffusion transformers (DiTs) have emerged as a dominant architecture for text-to-image generation, yet their performance drops when generating at resolutions beyond their training range. Existing training-free approaches mitigate this by modifying inference-time attention behavior, often through Rotary Position Embeddings (RoPE) extrapolation combined with attention scaling. However, these strategies apply a uniform and content-agnostic scaling across RoPE components with distinct frequency characteristics, inducing a trade-off between preserving global structure and recovering fine detail. We introduce SEGA, a training-free method that dynamically scales attention across RoPE components according to the latent's spatial-frequency structure at each denoising step. This adaptive scaling improves both structural coherence and fine-detail fidelity. Experiments show that SEGA consistently improves high-resolution synthesis across multiple target resolutions, outperforming state-of-the-art training-free baselines.
title SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2605.22668