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

TL;DR

SEGA is a training-free, adaptive attention scaling method for diffusion transformers that enhances high-resolution image synthesis by dynamically adjusting attention based on spatial-frequency content.

Contribution

SEGA introduces a novel, content-aware, frequency-guided attention scaling technique that improves resolution extrapolation without additional training.

Findings

SEGA outperforms existing training-free methods in high-resolution image generation.

It improves structural coherence and fine detail fidelity across multiple resolutions.

Experimental results demonstrate consistent performance gains over state-of-the-art baselines.

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.