CityU Hong Kong unveils RefineEdit, a training-free image editing framework that tops PIE-Bench
CityU-HongKong · hf · 2026-09-21
Researchers at City University of Hong Kong introduce RefineEdit, a training-free prompt-to-prompt image editing framework built on a Generative Refinement Network.
Key idea: couple edit localization with content generation by globally refining binary image codes. An editing branch is initialized from an intermediate source state (reusing the emerging layout), and the signed probability differences between the two branches over the same source-sampled bits select which positions and bits to edit — selected bits follow editing refinement while the rest copy the evolving source state.
Stabilization mechanisms:
- adaptive spatial freezing limits unnecessary mask expansion
- finite bit locking keeps recently selected bits editable
- no extra training, external masks, or attention control needed
Results: across nine editing categories on PIE-Bench, RefineEdit achieves the best background-preservation scores (PSNR, LPIPS, MSE, SSIM) plus the highest whole-image and edited-region CLIP scores, addressing both incomplete edits and accidental alteration of unrelated regions that plague diffusion and causal autoregressive editors.
More from Multimodal
- Seedance 2.5 generates a photorealistic 30s Korean winter travel ad, full prompt included — SimplyAnnisa · 2026-09-21
- Open-source 3DGS tool LichtFeld goes end-to-end: real-time SLAM reconstruction at 30 FPS — janusch_patas · 2026-09-21
- Punch Monkey AI film hits 50M+ views as creator shares full generation workflow — HashemGhaili · 2026-09-21
- Vaporous glitch art prompt template for AI image generation — LudovicCreator · 2026-09-21
- Fully AI-generated music video made with Seedance 2.5 shows pricey MV production fading — aitrendz_xyz · 2026-09-21
- Qwen-Image 2.1 LoRA training plagued by anatomy failures, community hunts for fixes — ResidentFrame4195 · 2026-09-21