Researchers from Hugging Face have published a new paper addressing the challenges of multi-reference image editing, a task that requires maintaining visual consistency across multiple reference images while ensuring overall harmony. The paper, titled "Evaluation-Verification Reward for Consistent Multi-Reference Image Editing," introduces a novel reward model called the Multi-dimensional Evaluation-Verification Reward (EVR).
The authors note that while reinforcement learning (RL) has been effective for text-to-image generation and single-image editing, its application to multi-reference editing has been limited by the lack of suitable reward models that can capture multi-image relational constraints. Naively using multimodal large language models (MLLMs) as zero-shot evaluators also faces a tension between hallucination-prone long-form reasoning and the limited deductive power of short-form judgments.
To overcome these issues, 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. This process produces reliable and fine-grained reward signals. The method is combined with a scalable data pipeline, enabling RL fine-tuning of off-the-shelf editors without architectural changes.
Extensive experiments demonstrate substantial gains over the base Qwen-Image-Edit model, with improvements in consistency and harmony that match or surpass the performance of NanoBanana, a state-of-the-art editing model. The paper is available on Hugging Face and has been shared via the platform's automated librarian bot, which also recommended several related papers on reinforcement learning and image editing.