PVC-Judge is a state-of-the-art 8B assessment model for evaluating image editing models in visual consistency.

🚀 Quick Start!

Clone github repo

git clone https://github.com/ZhangqiJiang07/GEditBench_v2.git
cd GEditBench_v2

Option 1: Packaged as an online client

• Merge LoRA weights to models, required env torch/peft/transformers

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

• Implement online server via vLLM

python -m vllm.entrypoints.openai.api_server \
  --model /path/to/save/PVC/Judge/model \
  --served-model-name PVC-Judge \
  --tensor-parallel-size 1 \
  --mm-encoder-tp-mode data \
  --limit-mm-per-prompt.video 0 \
  --host 0.0.0.0 \
  --port 25930 \
  --dtype bfloat16 \
  --gpu-memory-utilization 0.80 \
  --max_num_seqs 32 \
  --max-model-len 48000 \
  --distributed-executor-backend mp

• Use autopipeline for inference.

See our repo for detailed usage!

Option 2: Offline Inference

# For local judge inference
conda env create -f environments/pvc_judge.yml
conda activate pvc_judge
# or:
python3.12 -m venv .venvs/pvc_judge
source .venvs/pvc_judge/bin/activate
python -m pip install -r environments/requirements/pvc_judge.lock.txt


# Run
bash ./scripts/local_eval.sh vc_reward