PrefEdit: Fine-Grained Human-Centric Image Editing with Failure-Aware Rubric Rewards

Diffusion-based image editing has made remarkable progress in open-ended instruction following and broad semantic manipulation. Diffusion-based models perform well on broad semantic manipulation, but fine-grained human editing requires precise local changes while preserving non-target content. To address such limitations, we first construct $\text{Prehuman-55K}$ , a high quality, large-scale dataset covering 50 anatomically localized editing subtasks ($e.g.,$ eye makeup, hairline shape, double chin). We then present $\text{PrefEdit}$, which is specialized in fine-grained human image editing. Firstly, a lightweight LE-LoRA module stage learns preservation-sensitive local-editing priors. Based on such local editing ability, our core idea is to design a failure-aware rubric reward: Subtask-Specific Rubrics define task-dependent success and preservation criteria, while failure-conditioned rubric correction prevents critical errors from being misleading by unrelated high scores. Extensive in-domain and out-of-domain experiments demonstrate that PrefEdit achieves state-of-the-art performance in instruction fulfillment, localization, preservation, and visual quality while substantially reducing critical failures. The code will be available after accepted.