Evaluation-Verification Reward for Consistent Multi-Reference Image Editing
Yingmao Miao, Pengfei Zhang, Xiaochen Lv, Meng Yu, Lei Sun, Xiangxiang Chu, Chao Shen, Chenhao Lin
This paper develops a new method to evaluate and verify image editing consistency across multiple references, addressing a challenge in reinforcement learning for multi-reference editing. Practitioners may care about this approach as it enables more accurate and reliable reinforcement learning for image editing tasks.
Abstract
While recent image editing models have made rapid progress, multi-reference editing remains challenging, particularly in maintaining visual consistency across references and ensuring overall visual harmony. Reinforcement learning has proven highly effective for text-to-image generation and single-image editing, but its extension to multi-reference editing is hindered by the absence of suitable reward models that capture multi-image relational constraints. Moreover, naively using multimodal large language models(MLLMs) as zero-shot evaluators faces a key tension between hallucination-prone long-form reasoning and the limited deductive power of short-form judgments. We address these issues with a Multi-dimensional Evaluation-Verification Reward(EVR). EVR decomposes evaluation into distinct visual criteria; for each criterion, an MLLM Evaluator generates multiple candidate hypotheses, and a Verifier grounds each claim in concrete visual evidence to accept or reject it, producing reliable and fine-grained reward signals. Together with a scalable data pipeline, our method enables RL fine-tuning of off-the-shelf editors without architectural changes. Extensive experiments show substantial gains over the base Qwen-Image-Edit, improving consistency and harmony to match or surpass NanoBanana.