EditBridge: A Diffusion Bridge Framework for Faithful, Efficient Ultra-High-Resolution Image Editing

Forum topic · 小凯 · 2026-08-20

Summary

EditBridge (arXiv:2608.18063) is a diffusion bridge framework for ultra-high-resolution image editing. Existing diffusion-based editors are limited to sub-1K resolutions due to quadratic attention complexity and memory costs. The common workaround — low-resolution editing followed by independent super-resolution — suffers from information divergence (hallucinated details contradicting the original HR source) and texture degradation (over-smoothed or over-sharpened artifacts). EditBridge instead formulates refinement as a structured data-to-data transformation from the low-resolution edited result to its high-resolution counterpart, conditioned explicitly on the original HR source to preserve authentic details. A prior-guided tiled sparse attention mechanism exploits semantic correspondences from the first-stage edit to restrict cross-image interactions to spatially aligned regions, substantially cutting computational overhead. Experiments show faithful editing and superior perceptual quality at resolutions up to 4K, a 3.6-8.4x speedup at 2K, and practical 4K editing in 61 seconds.

Paper Overview

Field: Computer Vision
Authors: Jiayi Song, Shijie Huang, Fangtai Wu, Yubo Huang, Zhenxiong Tan, Songhua Liu, Jiaming Liu, Ruihua Huang
arXiv: 2608.18063

Abstract

High-resolution image editing is increasingly demanded in professional workflows, yet existing diffusion-based models remain constrained to resolutions below 1K due to quadratic attention complexity and prohibitive memory requirements. A prevalent workaround employs a two-stage pipeline: editing at low resolution followed by independent super-resolution. However, this approach suffers from two critical issues: information divergence, where hallucinated details contradict the original high-resolution (HR) source, and texture degradation, manifesting as over-smoothed or over-sharpened artifacts.

We propose EditBridge, a diffusion bridge framework for efficient ultra high-resolution editing. Unlike conventional diffusion that regenerates from noise, we formulate refinement as structured data-to-data transformation from the low-resolution (LR) edited result to its HR counterpart, explicitly conditioned on the original HR source to preserve authentic details. To efficiently incorporate HR source guidance, we introduce a prior-guided tiled sparse attention mechanism that leverages semantic correspondences from the first-stage edit to restrict cross-image interactions to spatially aligned regions, significantly reducing computational overhead.

Key Results

High-fidelity editing and superior perceptual quality at resolutions up to 4K
3.6-8.4x speedup at 2K resolution
Practical 4K editing in 61 seconds