Research map / Image generation / Super-resolution
EditBridge: Towards Faithful and Efficient Ultra-High-Resolution Image Editing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EditBridge, a diffusion-bridge refinement stage that upscales low-resolution edits to 2K/4K conditioned on the HR source, with attention-derived prior-guided block-wise sparse attention to cut cross-image cost.
Problem
High-resolution (up to 4K) instruction-based image editing with diffusion transformers is limited by quadratic attention cost, and two-stage edit-then-super-resolution pipelines hallucinate details that diverge from the HR source and degrade textures.
Approach
Formulates HR refinement as a Brownian-bridge data-to-data translation from the upsampled LR edit to the HR edit, conditioned on the HR source, and sparsifies target-source cross-attention to windows around correspondences derived from first-stage LR editing attention maps.
Key results
- At 2K, EditBridge runs in 4.08 s versus 14.82 s for DiT-SR and 113.32 s for DiT4SR, with mPSNR 18.64 and mLPIPS 0.390.
- At 4K, EditBridge reports mPSNR 24.14, mLPIPS 0.211 and 61.10 s total time, against 108.72 s for TSD-SR (mPSNR 22.28).
- At 1K, the method attains the best HaarPSI (0.499) and mPSNR (18.65) among compared methods, though its mLPIPS (0.383) is slightly behind ScaleEdit (0.378).
Surprising findings
- Single-step inference (N=1) outperforms 5 and 10 sampling steps on all metrics, which the authors attribute to quantization error accumulation at high resolution.
- Full attention scores higher on some pixel metrics (e.g., 1K HaarPSI 0.539 vs 0.499) yet shows ghosting artifacts, so the authors argue these metrics can be misleading for editing.
Limitations
- Relies on a pre-computed correspondence prior that must be extracted before translation, a bottleneck for fully automated pipelines.
- Iterative bridge sampling still incurs non-negligible overhead at high resolution.
- Single-step result and sparse-attention window size k are not thoroughly characterized in the main text.
Experimental setup
Datasets: Aesthetic-4k Aesthetic-Train-V2
Benchmarks: internal 1K/2K/4K editing test set following ScaleEdit protocol
Metrics: HaarPSI mPSNR mSSIM mMSE mLPIPS inference time
Baselines: Direct Inference DiT4SR DiT-SR PiSA-SR TSD-SR HiFlow ScaleEdit
Builds on: Qwen-Image-Edit VMoBA FlashAttention Prodigy LoRA ScaleEdit
Closest papers by meaning
- h-Edit: Effective and Flexible Diffusion-Based Editing via Doob's h-Transform (CVPR 2025 · similarity 0.88)
- UltraEdit: Instruction-based Fine-Grained Image Editing at Scale (NeurIPS 2024 · similarity 0.87)
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.87)
- PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.87)
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.87)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.