EditReward-Bench

Name: EditReward-Bench
Creator: maas
Published: 2025-10-04T11:32:47+08:00

魔搭社区2026-06-22 更新2025-10-04 收录

图像编辑

模型评估

数据链接：

https://modelscope.cn/datasets/TIGER-Lab/EditReward-Bench 数据链接链接失效反馈

官方服务：

资源简介：

# EditReward-Bench: A Human-Aligned Reward Model Benchmark for Instruction-Guided Image Editing [![Project Website](https://img.shields.io/badge/%F0%9F%8C%90-Project%20Website-deepgray)](https://tiger-ai-lab.github.io/EditReward/) [![arXiv](https://img.shields.io/badge/arXiv-2509.26346-b31b1b.svg)](https://arxiv.org/abs/2509.26346) [![Hugging Face Model](https://img.shields.io/badge/%F0%9F%A4%97-Model-yellow)](https://huggingface.co/collections/TIGER-Lab/editreward-68ddf026ef9eb1510458abc6) [![Hugging Face Dataset](https://img.shields.io/badge/%F0%9F%A4%97-Dataset-green)](https://huggingface.co/datasets/TIGER-Lab/EditReward-Data) [![GitHub](https://img.shields.io/badge/Code-GitHub-darkblue)](https://github.com/TIGER-AI-Lab/EditReward) This repository contains **EditReward-Bench**, a new benchmark introduced in the paper [EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing](https://huggingface.co/papers/2509.26346). ## Introduction Recent advances in image editing with natural language instructions have shown promising progress, particularly with closed-source models. However, open-source models often lag due to the lack of a reliable reward model for scaling up high-quality synthetic training data. To address this critical bottleneck, the paper introduces **EditReward**, a human-aligned reward model trained with a new large-scale human preference dataset called **EditReward-Data** (over 200K preference pairs). To rigorously assess **EditReward** and future models, we also introduce **EditReward-Bench**, a new benchmark built upon high-quality annotations, which includes more difficult multi-way preference prediction tasks. EditReward demonstrates superior alignment with human preferences in instruction-guided image editing tasks, achieving state-of-the-art human correlation on established benchmarks and on EditReward-Bench itself. * **Paper**: https://huggingface.co/papers/2509.26346 * **Project Page**: https://tiger-ai-lab.github.io/EditReward * **Code**: https://github.com/TIGER-AI-Lab/EditReward ### Dataset Statistics | Dataset | Original Samples | Pairwise Comparisons | File Size | |---------|-----------------|---------------------|-----------| | 2-pair | 133 | 133 | ~333 MB | | 3-pair | 150 | 450 | ~833 MB | | 4-pair | 148 | 888 | ~1.5 GB | | **Total** | **431** | **1471** | **~2.7 GB** | - **Total unique instructions**: 431 - **Score range**: 2.0 - 8.0 (Overall score = Instruction Following + Visual Quality) - **Average score difference**: 2.23 - **Total images**: 1,739 (all embedded in parquet files) ## Download EditReward-Bench You can download the **EditReward-Bench** dataset using the Hugging Face CLI: ```bash huggingface-cli download --repo-type dataset TIGER-Lab/EditReward-Bench --local-dir /your-local-dataset-path ``` ## Usage Please refer to the github code for evaluation! The EditReward-bench's evaluation is available at [EditReward/evaluate](https://github.com/TIGER-AI-Lab/EditReward/tree/main/EditReward/evaluate). ## Benchmark Results: Alignment with Humans The following tables present the experimental results from the paper, showcasing various models' alignment with human preferences on established benchmarks and on **EditReward-Bench**. <details open> <summary> Experimental Results: Alignment with Humans </summary> | Method | GenAI-Bench | AURORA-Bench | ImagenHub | EditReward-Bench (Overall) | | :--- | :--- | :--- | :--- | :--- | | Random | 25.90 | 33.43 | -- | 13.84 | | Human-to-Human | -- | -- | 41.84 | -- | | ***Proprietary Models*** | | | | | | GPT-4o | 53.54 | 50.81 | 38.21 | 28.31 | | GPT-5 | 59.61 | 47.27 | 40.85 | 37.81 | | Gemini-2.0-Flash | 53.32 | 44.31 | 23.69 | 33.47 | | Gemini-2.5-Flash | 57.01 | 47.63 | **41.62** | 38.02 | | ***Open-Source VLMs*** | | | | | | Qwen2.5-VL-3B-Inst | 42.76 | 30.69 | -2.54 | 26.86 | | Qwen2.5-VL-7B-Inst | 40.48 | 38.62 | 18.59 | 29.75 | | Qwen2.5-VL-32B-Inst | 39.28 | 37.06 | 26.87 | 28.72 | | MiMo-VL-7B-SFT-2508 | 57.89 | 30.43 | 22.14 | 31.19 | | ADIEE | 59.96 | 55.56 | 34.50 | -- | | ***Reward Models (Ours)*** | | | | | | EditReward (on Qwen2.5-VL-7B) | 63.97 | 59.50 | 36.18 | 36.78 | | EditReward (on MiMo-VL-7B) | **65.72** | **63.62** | 35.20 | **38.42** | </details> --- <details open> <summary> EditReward-Bench Results (K-way preference prediction) </summary> | Method | EditReward-Bench (K=2) | EditReward-Bench (K=3) | EditReward-Bench (K=4) | EditReward-Bench (Overall) | | :--- | :--- | :--- | :--- | :--- | | Random | 25.81 | 11.33 | 1.35 | 13.84 | | Human-to-Human | -- | -- | -- | -- | | ***Proprietary Models*** | | | | | | GPT-4o | 45.69 | 27.33 | 7.31 | 28.31 | | GPT-5 | 57.53 | 38.51 | 12.84 | 37.81 | | Gemini-2.0-Flash | 52.43 | 33.33 | **13.51** | 33.47 | | Gemini-2.5-Flash | **58.61** | 39.86 | 12.16 | 38.02 | | ***Open-Source VLMs*** | | | | | | Qwen2.5-VL-3B-Inst | 51.07 | 20.27 | 2.71 | 26.86 | | Qwen2.5-VL-7B-Inst | 52.69 | 24.67 | 3.38 | 29.75 | | Qwen2.5-VL-32B-Inst | 50.54 | 25.27 | 4.05 | 28.72 | | MiMo-VL-7B-SFT-2508 | 49.46 | 30.41 | 9.46 | 31.19 | | ADIEE | -- | -- | -- | -- | | ***Reward Models (Ours)*** | | | | | | EditReward (on Qwen2.5-VL-7B) | 56.99 | 36.00 | 10.81 | 36.78 | | EditReward (on MiMo-VL-7B) | 56.45 | **42.67** | 11.49 | **38.42** | </details> ## Citation Please kindly cite our paper if you use our code, data, models or results: ```bibtex @article{wu2025editreward, title={EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing}, author={Wu, Keming and Jiang, Sicong and Ku, Max and Nie, Ping and Liu, Minghao and Chen, Wenhu}, journal={arXiv preprint arXiv:2509.26346}, year={2025} } ```

<img src="assets/logo.png" width="65%"> <a href="https://vectorspacelab.github.io/EditScore"><img src="https://img.shields.io/badge/Project%20Page-EditScore-yellow" alt="项目主页"></a> <a href="https://arxiv.org/abs/2509.23909"><img src="https://img.shields.io/badge/arXiv%20paper-2509.23909-b31b1b.svg" alt="arXiv论文"></a> <a href="https://huggingface.co/collections/EditScore/editscore-68d8e27ee676981221db3cfe"><img src="https://img.shields.io/badge/EditScore-🤗-yellow" alt="模型"></a> <a href="https://huggingface.co/datasets/EditScore/EditReward-Bench"><img src="https://img.shields.io/badge/EditReward--Bench-🤗-yellow" alt="数据集"></a> <h4 align="center"> <a href="#-最新动态">最新动态</a> | <a href="#-快速上手">快速上手</a> | <a href="#-基准测试你的图像编辑奖励模型">基准测试使用</a> | <a href="#-引用本文">引用本文</a> </h4> **EditScore** 是一系列前沿开源奖励模型（7B–72B参数），旨在评估并优化指令引导的图像编辑。 ## ✨ 核心亮点 - **最先进性能表现**：可媲美头部闭源视觉语言模型（Visual Language Model, VLM）。通过自集成策略，我们的最大尺寸模型甚至在综合基准测试集**EditReward-Bench**上超越了GPT-5。 - **可靠的评估标准**：我们推出了**EditReward-Bench**，这是首个专为图像编辑领域奖励模型评估设计的公开基准测试集，涵盖13个子任务、11款最先进的编辑模型（包含闭源模型）以及专家人工标注。 - **简洁易用**：仅需数行代码即可为你的图像编辑结果获取精准的质量评分。 - **多场景适配**：可作为顶尖的重排序器优化编辑输出，也可作为高保真奖励信号用于**强化学习（Reinforcement Learning, RL）** 微调，实现稳定且高效的训练。 ## 🔥 最新动态 - **2025-09-30**：我们发布了**OmniGen2-EditScore7B**，通过高保真EditScore解锁图像编辑的在线强化学习。其低秩自适应（Low-Rank Adaptation, LoRA）权重可在[Hugging Face](https://huggingface.co/OmniGen2/OmniGen2-EditScore7B)和[ModelScope](https://www.modelscope.cn/models/OmniGen2/OmniGen2-EditScore7B)获取。 - **2025-09-30**：我们正式发布**EditScore**与**EditReward-Bench**！模型权重与基准数据集现已公开可用，可通过以下平台获取：Hugging Face的[模型合集](https://huggingface.co/collections/EditScore/editscore-68d8e27ee676981221db3cfe)与[基准数据集](https://huggingface.co/datasets/EditScore/EditReward-Bench)，以及ModelScope的[模型合集](https://www.modelscope.cn/collections/EditScore-8b0d53aa945d4e)与[基准数据集](https://www.modelscope.cn/datasets/EditScore/EditReward-Bench)。 ## 📖 项目简介强化学习（Reinforcement Learning, RL）在该领域拥有巨大潜力，但其进展却因缺乏高保真、高效的奖励信号而受到严重阻碍。为突破这一瓶颈，我们提供了一套系统性的两阶段解决方案： - **严谨的评估标准**：我们首先推出**EditReward-Bench**，一款用于直接且可靠评估奖励模型的全新公开基准测试集。该基准涵盖13个多样化子任务与专家人工标注，为衡量奖励信号质量确立了黄金标准。 - **强大且通用的工具**：基于该基准的指导，我们开发了**EditScore**模型系列。通过精心的数据筛选与高效的自集成策略，EditScore将开源奖励模型的性能推至新高度，其精度甚至超越了领先的闭源视觉语言模型。 <img src="assets/table_reward_model_results.png" width="95%"> EditReward-Bench上的基准测试结果。 我们通过两个关键应用场景展示了EditScore的实用价值： - **作为顶尖重排序器**：使用EditScore进行N选最优选择，可快速提升各类编辑模型的输出质量。 - **作为强化学习的高保真奖励信号**：使用EditScore作为稳健的奖励信号，通过强化学习对模型进行微调，实现稳定训练并解锁显著的性能提升，而通用视觉语言模型在此类任务中往往难以奏效。本仓库同时发布**EditScore**模型与**EditReward-Bench**数据集，以推动奖励建模、策略优化以及AI驱动的模型改进领域的未来研究。 <img src="assets/figure_edit_results.png" width="95%"> EditScore作为图像编辑的优质奖励信号。 ## 📌 后续计划我们正在积极优化EditScore并拓展其功能，后续计划包括： - [ ] 发布将EditScore应用于OmniGen2的强化学习训练代码。 - [ ] 为OmniGen2、Flux-dev-Kontext以及Qwen-Image-Edit提供N选最优推理脚本。 ## 🚀 快速上手 ### 🛠️ 环境搭建 #### ✅ 标准安装流程 bash # 1. 克隆仓库 git clone git@github.com:VectorSpaceLab/EditScore.git cd EditScore # 2.（可选）创建干净的Python环境 conda create -n editscore python=3.12 conda activate editscore # 3. 安装依赖 # 3.1 安装PyTorch（请匹配对应CUDA版本） pip install torch==2.7.1 torchvision --extra-index-url https://download.pytorch.org/whl/cu126 # 3.2 安装其他依赖包 pip install -r requirements.txt # 即使不安装vllm，EditScore也可正常运行，但推荐安装以获得最佳性能。 pip install vllm #### 🌏 中国大陆用户加速方案 bash # 从国内镜像安装PyTorch pip install torch==2.7.1 torchvision --index-url https://mirror.sjtu.edu.cn/pytorch-wheels/cu126 # 从清华镜像安装其他依赖 pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple # 即使不安装vllm，EditScore也可正常运行，但推荐安装以获得最佳性能。 pip install vllm -i https://pypi.tuna.tsinghua.edu.cn/simple --- ### 🧪 代码示例使用EditScore非常简便。模型将在首次运行时自动从Hugging Face Hub下载。 python from PIL import Image from editscore import EditScore # 加载EditScore模型，首次运行时将自动下载。 # 替换为你想要使用的具体模型版本。 model_path = "Qwen/Qwen2.5-VL-7B-Instruct" lora_path = "EditScore/EditScore-7B" scorer = EditScore( backbone="qwen25vl", # 若需更快推理，可设置为"qwen25vl_vllm" model_name_or_path=model_path, enable_lora=True, lora_path=lora_path, score_range=25, num_pass=1, # 增加该参数可通过自集成策略提升性能 ) input_image = Image.open("example_images/input.png") output_image = Image.open("example_images/output.png") instruction = "将背景调整为玻璃幕墙。" result = scorer.evaluate([input_image, output_image], instruction) print(f"编辑评分：{result['final_score']}") # 预期输出：包含最终评分及其他细节的字典。 --- ## 📊 基准测试你的图像编辑奖励模型我们提供了一套评估脚本，用于在**EditReward-Bench**上对奖励模型进行基准测试。若要评估自定义的奖励模型，仅需创建一个具有相似接口的评分器类，并更新该脚本即可。 bash # 该脚本将在基准测试集上评估默认的EditScore模型 bash evaluate.sh # 若需使用VLLM加速推理，可执行： bash evaluate_vllm.sh ## ❤️ 引用本文若本仓库或您的工作对您有所帮助，欢迎点亮⭐并引用我们的论文，感谢您的支持： bibtex @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} }

提供机构：

maas

创建时间：

2025-10-04

搜集汇总

数据集介绍