EDITBRIDGE: Towards Faithful and Efficient Ultra-High-Resolution Image Editing

The core problem EditBridge solves is embarrassingly familiar to anyone who has tried to get a diffusion model to edit a 4K photograph: you either run out of memory, wait an eternity, or accept that the super-resolution step will invent details that contradict the original image. The two-stage workaround—edit at low resolution, upscale independently—fails because the SR model has never seen the original high-resolution source. It hallucinates. Faces change. Doors become unrecognizable. Textures either smear or sharpen into artifacts.

EditBridge reframes the second stage entirely. Rather than treating upscaling as a generation problem starting from noise, it treats it as a structured translation between two data distributions: the upsampled low-resolution edit on one end, the desired high-resolution output on the other. This is the diffusion bridge formulation—Brownian bridge stochastic interpolation between paired endpoints—and it matters because the structural information from the low-resolution edit is preserved throughout the refinement trajectory rather than discarded at the start of a fresh denoising chain.

The more interesting engineering contribution is the Prior-Guided Block-wise Sparse Attention mechanism. Naively concatenating high-resolution source tokens into a Diffusion Transformer's attention sequence is computationally ruinous—token count scales quadratically with resolution. The insight here is that not every target patch needs to attend to every source patch. For unedited regions, fidelity comes from attending to spatially corresponding anchors in the source. For edited regions, local context suffices. EditBridge extracts a correspondence map from the attention matrices of the first-stage low-resolution edit, upscales those correspondences to the target resolution via nearest-neighbor expansion, and uses them to constrain cross-domain attention to a local window around each semantically aligned anchor. This reduces cross-domain attention complexity from quadratic to linear in the number of tokens.

The ablation on attention type is worth dwelling on. Full attention actually scores slightly higher on pixel-level reconstruction metrics like PSNR. But the paper argues—and the qualitative figures support—that this is misleading: global attention causes the model to bleed information from semantically irrelevant source regions into newly synthesized areas, producing ghosting and structural artifacts in edited zones. Sparse attention trades a marginal metric advantage for visually cleaner edits. That tradeoff is defensible.

On inference speed, the numbers are concrete. At 2K resolution, EditBridge runs in roughly 4 seconds, representing a 3.6 to 8.4 times speedup over diffusion-based SR baselines. At 4K, the speedup over SR methods narrows to 1.8 to 2.1 times, but the comparison against direct high-resolution inference shows an 11-times acceleration, completing in 61 seconds. Single-step inference turns out to be optimal—more steps accumulate quantization errors at extreme resolutions rather than improving quality.

The stated limitations are honest. The correspondence prior must be extracted before refinement begins, which creates a dependency on first-stage quality and complicates fully automated pipelines. The authors acknowledge this and propose learnable prior estimation as future work. The framework is also built on top of Qwen-Image-Edit as its first-stage model, so its ceiling is partly determined by that base.