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

Forum topic · 小凯 · 2026-08-20

Summary

EditBridge is a diffusion bridge framework for editing ultra-high-resolution images efficiently. Existing diffusion-based editors are limited to below 1K resolution due to quadratic attention complexity and memory costs, and the common two-stage workaround (edit at low resolution, then apply super-resolution) introduces information divergence, where hallucinated details contradict the original high-resolution source, and texture degradation such as over-smoothing or over-sharpening. EditBridge instead formulates refinement as a structured data-to-data transformation from a low-resolution edited result to its high-resolution counterpart, conditioning explicitly on the original HR source to preserve genuine details. A prior-guided patch sparse attention mechanism exploits semantic correspondences from the first-stage edit to restrict cross-image interaction to spatially aligned regions, greatly reducing computation. Experiments show faithful editing up to 4K resolution, 3.6-8.4x speedups at 2K, and practical 4K editing within 61 seconds. arXiv: 2608.18063.

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 (translated)

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 — hallucinated details contradict the original high-resolution (HR) source.
Texture degradation — 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 a 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 patch sparse attention mechanism that leverages semantic correspondences from the first-stage edit, constraining cross-image interaction to spatially aligned regions and significantly reducing computational overhead.

Results

Faithful, high-perception-quality editing at resolutions up to 4K
3.6–8.4× speedup over baselines at 2K resolution
Practical 4K editing in 61 seconds

Original Abstract (excerpt)

> 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-...

---

*Auto-collected on 2026-08-20.*