license: apache-2.0
language:
- en
task_categories:
- image-text-to-text
tags:
- image-editing
- reward-modeling
- reinforcement-learning
- benchmark
- evaluation
- multimodal
News | Quick Start | Benchmark Usage | Citation
EditScore is a series of state-of-the-art open-source reward models (7B–72B) designed to evaluate and enhance instruction-guided image editing.
Paper Abstract
Instruction-guided image editing has achieved remarkable progress, yet current models still face challenges with complex instructions and often require multiple samples to produce a desired result. Reinforcement Learning (RL) offers a promising solution, but its adoption in image editing has been severely hindered by the lack of a high-fidelity, efficient reward signal. In this work, we present a comprehensive methodology to overcome this barrier, centered on the development of a state-of-the-art, specialized reward model. We first introduce EditReward-Bench, a comprehensive benchmark to systematically evaluate reward models on editing quality. Building on this benchmark, we develop EditScore, a series of reward models (7B-72B) for evaluating the quality of instruction-guided image editing. Through meticulous data curation and filtering, EditScore effectively matches the performance of learning proprietary VLMs. Furthermore, coupled with an effective self-ensemble strategy tailored for the generative nature of EditScore, our largest variant even surpasses GPT-5 in the benchmark. We then demonstrate that a high-fidelity reward model is the key to unlocking online RL for image editing. Our experiments show that, while even the largest open-source VLMs fail to provide an effective learning signal, EditScore enables efficient and robust policy optimization. Applying our framework to a strong base model, OmniGen2, results in a final model that shows a substantial and consistent performance uplift. Overall, this work provides the first systematic path from benchmarking to reward modeling to RL training in image editing, showing that a high-fidelity, domain-specialized reward model is the key to unlocking the full potential of RL in this domain.
✨ Highlights
- State-of-the-Art Performance: Effectively matches the performance of leading proprietary VLMs. With a self-ensembling strategy, our largest model surpasses even GPT-5 on our comprehensive benchmark, EditReward-Bench.
- A Reliable Evaluation Standard: We introduce EditReward-Bench, the first public benchmark specifically designed for evaluating reward models in image editing, featuring 13 subtasks, 11 state-of-the-art editing models (including proprietary models) and expert human annotations.
- Simple and Easy-to-Use: Get an accurate quality score for your image edits with just a few lines of code.
- Versatile Applications: Ready to use as a best-in-class reranker to improve editing outputs, or as a high-fidelity reward signal for stable and effective Reinforcement Learning (RL) fine-tuning.
🔥 News
- 2025-09-30: We release OmniGen2-EditScore7B, unlocking online RL For Image Editing via high-fidelity EditScore. LoRA weights are available at Hugging Face and ModelScope.
- 2025-09-30: We are excited to release EditScore and EditReward-Bench! Model weights and the benchmark dataset are now publicly available. You can access them on Hugging Face: Models Collection and Benchmark Dataset, and on ModelScope: Models Collection and Benchmark Dataset.
📖 Introduction
While Reinforcement Learning (RL) holds immense potential for this domain, its progress has been severely hindered by the absence of a high-fidelity, efficient reward signal.
To overcome this barrier, we provide a systematic, two-part solution:
A Rigorous Evaluation Standard: We first introduce EditReward-Bench, a new public benchmark for the direct and reliable evaluation of reward models. It features 13 diverse subtasks and expert human annotations, establishing a gold standard for measuring reward signal quality.
A Powerful & Versatile Tool: Guided by our benchmark, we developed the EditScore model series. Through meticulous data curation and an effective self-ensembling strategy, EditScore sets a new state of the art for open-source reward models, even surpassing the accuracy of leading proprietary VLMs.
Benchmark results on EditReward-Bench.
We demonstrate the practical utility of EditScore through two key applications:
- As a State-of-the-Art Reranker: Use EditScore to perform Best-of-N selection and instantly improve the output quality of diverse editing models.
- As a High-Fidelity Reward for RL: Use EditScore as a robust reward signal to fine-tune models via RL, enabling stable training and unlocking significant performance gains where general-purpose VLMs fail.
This repository releases both the EditScore models and the EditReward-Bench dataset to facilitate future research in reward modeling, policy optimization, and AI-driven model improvement.
EditScore as a superior reward signal for image editing.
📌 TODO
We are actively working on improving EditScore and expanding its capabilities. Here's what's next:
- Release RL training code applying EditScore to OmniGen2.
- Provide Best-of-N inference scripts for OmniGen2, Flux-dev-Kontext, and Qwen-Image-Edit.
🚀 Quick Start
🛠️ Environment Setup
✅ Recommended Setup
# 1. Clone the repo
git clone git@github.com:VectorSpaceLab/EditScore.git
cd EditScore
# 2. (Optional) Create a clean Python environment
conda create -n editscore python=3.12
conda activate editscore
# 3. Install dependencies
# 3.1 Install PyTorch (choose correct CUDA version)
pip install torch==2.7.1 torchvision --extra-index-url https://download.pytorch.org/whl/cu126
# 3.2 Install other required packages
pip install -r requirements.txt
# EditScore runs even without vllm, though we recommend install it for best performance.
pip install vllm
🌏 For users in Mainland China
# Install PyTorch from a domestic mirror
pip install torch==2.7.1 torchvision --index-url https://mirror.sjtu.edu.cn/pytorch-wheels/cu126
# Install other dependencies from Tsinghua mirror
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
# EditScore runs even without vllm, though we recommend install it for best performance.
pip install vllm -i https://pypi.tuna.tsinghua.edu.cn/simple
🧪 Usage Example
Using EditScore is straightforward. The model will be automatically downloaded from the Hugging Face Hub on its first run.
from PIL import Image
from editscore import EditScore
# Load the EditScore model. It will be downloaded automatically.
# Replace with the specific model version you want to use.
model_path = "Qwen/Qwen2.5-VL-7B-Instruct"
lora_path = "EditScore/EditScore-7B"
scorer = EditScore(
backbone="qwen25vl", # set to "qwen25vl_vllm" for faster inference
model_name_or_path=model_path,
enable_lora=True,
lora_path=lora_path,
score_range=25,
num_pass=1, # Increase for better performance via self-ensembling
)
input_image = Image.open("example_images/input.png")
output_image = Image.open("example_images/output.png")
instruction = "Adjust the background to a glass wall."
result = scorer.evaluate([input_image, output_image], instruction)
print(f"Edit Score: {result['final_score']}")
# Expected output: A dictionary containing the final score and other details.
📊 Benchmark Your Image-Editing Reward Model
We provide an evaluation script to benchmark reward models on EditReward-Bench. To evaluate your own custom reward model, simply create a scorer class with a similar interface and update the script.
# This script will evaluate the default EditScore model on the benchmark
bash evaluate.sh
# Or speed up inference with VLLM
bash evaluate_vllm.sh
❤️ Citing Us
If you find this repository or our work useful, please consider giving a star ⭐ and citation 🦖, which would be greatly appreciated:
@article{luo2025editscore,
title={EditScore: Unlocking Online RL for Image Editing via High-Fidelity Reward Modeling},
author={Xin Luo and Jiahao Wang and Chenyuan Wu and Shitao Xiao and Xiyan Jiang and Defu Lian and Jiajun Zhang and Dong Liu and Zheng Liu},
journal={arXiv preprint arXiv:2509.23909},
year={2025}
}