--- base_model: Qwen/Qwen3-VL-8B-Instruct license: mit library_name: peft pipeline_tag: image-text-to-text --- # PVC-Judge: Pairwise Visual Consistency Judge PVC-Judge is a state-of-the-art 8B assessment model for evaluating image editing models in visual consistency. It is a pairwise preference model designed to capture the preservation of identity, structure, and semantic coherence between edited and original images. The model was introduced in the paper [GEditBench v2: A Human-Aligned Benchmark for General Image Editing](https://arxiv.org/abs/2603.28547) and is implemented as a LoRA adapter for [Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct).
## 🚀 Quick Start To use PVC-Judge, you typically need to merge the LoRA weights with the base model. ### 1. Merge LoRA weights This step requires `torch`, `peft`, and `transformers`. ```bash python ./scripts/merge_lora.py \ --base-model-path /path/to/Qwen3/VL/8B/Instruct \ --lora-weights-path /path/to/LoRA/Weights \ --model-save-dir /path/to/save/PVC/Judge/model ``` ### 2. Deployment or Local Inference You can serve the merged model via vLLM or run local evaluation as described in the [official repository](https://github.com/ZhangqiJiang07/GEditBench_v2). **Local Inference:** ```bash # Setup environment conda env create -f environments/pvc_judge.yml conda activate pvc_judge # Run evaluation bash ./scripts/local_eval.sh vc_reward ``` ## Citation ```bibtex @article{jiang2026geditbenchv2, title={GEditBench v2: A Human-Aligned Benchmark for General Image Editing}, author={Zhangqi Jiang and Zheng Sun and Xianfang Zeng and Yufeng Yang and Xuanyang Zhang and Yongliang Wu and Wei Cheng and Gang Yu and Xu Yang and Bihan Wen}, journal={arXiv preprint arXiv:2603.28547}, year={2026} } ```