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

The Problem: High-Resolution Editing Breaks Down

Current image editing tools powered by diffusion models hit a wall around 1K resolution. The bottleneck is attention—a core mechanism in diffusion models that requires quadratic memory and compute, making 4K or 8K editing infeasible. The workaround is a simple two-stage approach: edit the image at low resolution, then upscale it independently. But this creates two failure modes. First, the upscaler hallucinates details that contradict what was actually in the original high-resolution image. Second, the texture warps—becoming either over-smoothed or over-sharpened. Neither problem is easy to fix because the upscaler never sees the original.

EditBridge: Conditioning on Ground Truth

EditBridge flips the problem. Instead of treating editing and upscaling as separate tasks, it conditions the entire refinement process on the original high-resolution source image. The pipeline works in three stages:

1. Edit at low resolution: Use a standard diffusion model to produce an edited low-resolution image.
2. Reformulate as translation: Rather than generating from scratch, treat the task as translating the edited low-res result into its high-resolution counterpart—a constrained problem that preserves details from the original.
3. Guide with sparse attention: To make this efficient, the authors introduce a prior-guided block-wise sparse attention mechanism. Instead of allowing every pixel to attend to every other pixel (quadratic complexity), the mechanism uses semantic correspondences learned during the first-stage editing to align blocks between the low-res edited image and the original high-res source. This means attention is only computed within spatially aligned regions, slashing memory and compute.

The key insight is that by anchoring refinement to the authentic original, hallucinations are suppressed and real texture is preserved.

Results and Practical Impact

The speedups are substantial. At 2K resolution, EditBridge is 3.6× to 8.4× faster than baseline approaches. More critically, it makes 4K editing practical—completing a full 4K edit in 61 seconds on a single GPU, something that was infeasible before. The authors evaluate across multiple datasets and show improvements in perceptual quality metrics and user studies, with faithful preservation of original details while respecting the user's edits.

The technique generalizes across different editing tasks (inpainting, local editing, style transfer) because the core idea is task-agnostic: condition on ground truth and use sparse attention to stay efficient.

Why It Matters

Professional workflows increasingly demand high-resolution editing. Photographers and designers working at 4K or higher can't wait 10+ minutes per edit or accept hallucinated artifacts. EditBridge closes a real gap—it enables fast, faithful editing at resolutions that were previously off-limits. The sparse attention mechanism is also a reusable primitive for other diffusion-based tasks that require high-resolution input guidance.