Paper Overview
Field: Computer Vision (CV) Authors: Ritesh Thawkar, Shubham Patle, Shravan Venkatraman, Rao Muhammad Anwer Published: 2026-10-08 arXiv: 2610.12469
Summary
Instruction-guided image editors have become highly capable, yet improving them further still depends on human-edited training pairs or external reward models. Such supervision is costly to obtain and can reward plausible failures: a realistic output may leave the requested change undone or alter content that should be preserved.
This work aims to improve a pretrained image editor using only its own generations, without human-edited targets or an external training-time reward model. The authors propose a self-evolving framework, Rubric-CEPR, which verifies the editor's own samples with its internal representations through a rubric-augmented Contrastive Edit-Preservation Reward (CEPR):
- A Planner proposes structured edit instructions from unlabeled images.
- The Editor samples multiple candidate edits.
- A frozen Critic scores each candidate with the editor's already-exposed features against decomposed rubric items: edit realization, removal of the old state, and content preservation.
- A non-compensatory gating mechanism rejects infeasible candidates.
- The best verified candidates are distilled back into the editor via lightweight adapters.
Results
On Qwen-Image-Edit, Rubric-CEPR improves ImgEdit from 4.36 to 4.60 (+5.5%), with object isolation improving by +24.9%; the gains transfer to GEdit-Bench and Complex-Edit. The same pipeline improves ImgEdit by +7.8% on Step1X-Edit.
---
*Auto-collected on 2026-10-10.*