Hugging Face Daily PapersYulong Chen, Ziqian Zhang, Haoyu Zhang1 min readpaperadvanced
Refinement Is Inherently Editable: Training-Free Prompt-to-Prompt Image Editing with Generative Refinement Network
Summary
RefineEdit is a training‑free framework that edits images by refining binary image codes through a generative refinement network, using probability differences to pick edit locations. It outperforms prior methods on background preservation and CLIP scores across nine editing categories without extra training or masks.
- RefineEdit selects editable bits by comparing signed probability differences of binary image codes between source and edit branches.
- Adaptive spatial freezing limits mask growth and finite bit locking keeps recently edited bits mutable, stabilizing refinement.
- The approach requires no extra training, external masks, or attention control, leveraging existing generative models.
- On PIE‑Bench, RefineEdit leads in background‑preservation metrics (PSNR, LPIPS, MSE, SSIM) and whole‑image/edited‑region CLIP scores.
Image‑editing tool developers and diffusion model researchers should care because it provides a precise, training‑free way to perform prompt‑guided edits while preserving unrelated content.
7/10