Rubric-CEPR
A self-evolving image editing framework that builds reward signals using the editor's own internal representations together with structured rubrics. It filters candidates through non-compensatory gating and performs self-distillation with a lightweight adapter, requiring neither human annotation nor an external reward model.
Timeline (last 90 days)
Rubric-CEPR is an image editing framework that can self-evolve using only samples generated by the editor itself, requiring no human targets and no external reward model during training.
Rubric-CEPR's loop consists of three roles: Planner, Editor, Frozen Critic.
Contrastive Edit-Preservation Reward (CEPR) decomposes scoring into multiple rubric check items: whether the edit is truly achieved, whether the old state is removed, and whether the content is preserved, with each dimension evaluated independently rather than merged into a total score.
Non-compensatory gates directly reject infeasible candidates, preventing 'partial success' from being mistakenly treated as a high-quality sample.
Rubric-CEPR distills the validated best candidate back into the editor through a lightweight adapter, updating only a small number of parameters rather than retraining the entire model.
On Qwen-Image-Edit, Rubric-CEPR improved ImgEdit from 4.36 to 4.60, a gain of +5.5%.
On Qwen-Image-Edit, Rubric-CEPR achieved a +24.9% improvement on the object isolation task.
Rubric-CEPR's method was transferred to two benchmarks, GEdit-Bench and Complex-Edit.
When the same pipeline was applied to Step1X-Edit, a +7.8% improvement was achieved on ImgEdit.
Since two editors with different architectures (Qwen-Image-Edit and Step1X-Edit) both benefit from Rubric-CEPR, this shows that the framework is decoupled from the specific model and possesses a certain generality.
1 more timeline events
All Facts (11)
Since two editors with different architectures (Qwen-Image-Edit and Step1X-Edit) both benefit from Rubric-CEPR, this shows that the framework is decoupled from the specific model and possesses a certain generality.
50%UnverifiedRubric-CEPR's loop consists of three roles: Planner, Editor, Frozen Critic.
50%UnverifiedContrastive Edit-Preservation Reward (CEPR) decomposes scoring into multiple rubric check items: whether the edit is truly achieved, whether the old state is removed, and whether the content is preserved, with each dimension evaluated independently rather than merged into a total score.
50%UnverifiedNon-compensatory gates directly reject infeasible candidates, preventing 'partial success' from being mistakenly treated as a high-quality sample.
50%UnverifiedRubric-CEPR distills the validated best candidate back into the editor through a lightweight adapter, updating only a small number of parameters rather than retraining the entire model.
50%UnverifiedOn Qwen-Image-Edit, Rubric-CEPR improved ImgEdit from 4.36 to 4.60, a gain of +5.5%.
50%UnverifiedOn Qwen-Image-Edit, Rubric-CEPR achieved a +24.9% improvement on the object isolation task.
50%UnverifiedRubric-CEPR's method was transferred to two benchmarks, GEdit-Bench and Complex-Edit.
50%UnverifiedWhen the same pipeline was applied to Step1X-Edit, a +7.8% improvement was achieved on ImgEdit.
50%UnverifiedThe authors have made Rubric-CEPR's code publicly available.
50%UnverifiedRubric-CEPR is an image editing framework that can self-evolve using only samples generated by the editor itself, requiring no human targets and no external reward model during training.
50%