Towards Faithful and EfficientUltra-High-Resolution Image Editing

Jiayi Song^1,2Shijie Huang²Fangtai Wu²Yubo Huang²Zhenxiong Tan³Songhua Liu^1,*Jiaming Liu^2,†Ruihua Huang²

¹School of Artificial Intelligence, Shanghai Jiao Tong University²Qwen Business Unit of Alibaba³National University of Singapore

^*Corresponding author^†Project lead

liusonghua@sjtu.edu.cn

Read paper Code

4Kediting resolution

61.1s4K inference

3.6–8.4×2K speedup

1 stepbridge refinement

Why EditBridge

High-resolution editing without losing the source.

Challenge

Existing diffusion editors are usually limited to sub-1K outputs. A low-resolution edit followed by independent super-resolution can hallucinate details or damage texture because the original high-resolution source is no longer available to guide refinement.

Our approach

EditBridge reframes refinement as a structured data-to-data diffusion bridge from the low-resolution edited result to its high-resolution counterpart. The original high-resolution image remains an explicit condition, while prior-guided block-wise sparse attention preserves aligned details without the cost of dense global attention.

Traditional diffusion and EditBridge pipelines, with examples of hallucinated details and texture preservation. — Motivation. EditBridge conditions refinement on the original HR source, preventing information divergence and texture degradation.

Method

A bridge from coarse edits to faithful 4K results.

A diffusion bridge transports the upsampled LR edit toward the HR target. Correspondence priors from the first-stage editor route each target chunk to semantically aligned source blocks.

EditBridge framework with diffusion bridge refinement, correspondence prior, and block-wise sparse attention. — Overview of EditBridge. PG-BSA combines intra-domain self-attention with prior-guided cross-chunk sparse attention.

Coarse edit

A pretrained editor first produces an instruction-aligned result at its native resolution.

Diffusion bridge

The upsampled edit becomes a structured starting point instead of regenerating the image from noise.

Sparse HR guidance

Semantic correspondence selects only the source blocks needed to recover faithful high-frequency detail.

Results

Sharper details, stronger fidelity, practical latency.

EditBridge consistently leads reconstruction and perceptual metrics from 1K through 4K while remaining substantially faster than direct inference and conventional diffusion-based super-resolution.

Qualitative 1K comparison across EditBridge and baseline super-resolution methods. — Qualitative comparison at 1K. Ours recovers fine details while remaining faithful to the HR source.

Qualitative 2K comparison across EditBridge and baseline super-resolution methods. — Qualitative comparison at 2K. EditBridge preserves text, object structure, and local texture.

Analysis

One step is enough.

Visual comparison between full attention and EditBridge sparse attention. — Attention sparsity. Full attention introduces source-induced artifacts; PG-BSA preserves alignment.

Comparison of one, five, and ten bridge inference steps. — Sampling steps. Single-step inference gives the best fidelity–efficiency balance.