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

Javad Rajabi, Kimia Shaban, Koorosh Roohi, David B. Lindell, Babak Taati

arXiv:2605.22668 · 2026-05-20 공개 · arXiv · PDF

training-free diffusion-transformers rope-extrapolation spectral-energy resolution-extrapolation attention-scaling frequency-structure high-resolution-synthesis

Abstract

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.

한국어 요약

한 줄 요약

**[Diffusion Transformer / Resolution Extrapolation]** SEGA가 RoPE component별로 latent의 spatial-frequency 구조에 따라 동적 scaling -- training-free·multi-resolution에서 structural coherence·fine-detail fidelity 동시 개선·SOTA training-free baseline outperform.

핵심 기여도

핵심 아이디어

DiT의 해상도 외삽 실패는 RoPE 컴포넌트의 frequency 특성을 무시하는 균일 scaling에서 비롯되며, latent의 spatial-frequency 구조에 따라 컴포넌트별로 동적 scaling을 적용하면 training 없이도 structural coherence와 fine detail을 동시 회복할 수 있다.

기술적 접근법

주요 결과

의의 및 한계

**의의**: DiT의 해상도 외삽 한계를 spectral-energy 가이드 attention으로 training 없이 해결, RoPE 컴포넌트의 frequency 특성을 첫 명시적 활용, 다중 해상도 일관 성능. **한계**: 매우 극단적 해상도(8K 이상)에서의 검증 한계, latent spatial-frequency 분석의 계산 비용, abstract에서 quantitative 수치 미명시.

실용적 활용