EditHF-1M is a million-scale image editing dataset containing over 29M human preference pairs and 148K human mean opinion scores (MOS), evaluated across three dimensions: visual quality, editing alignment, and attribute preservation.
IEQA: A subset of the EditHF-1M dataset is adopted as the IEQA dataset for the New Trends in Image Restoration and Enhancement (NTIRE) Workshop and Challenge @ CVPR 2026, under the X-AIGC Quality Assessment – Track 2: Image Editing.
You can download the IEQA dataset from the following link: IEQA
EditHF is an MLLM-based evaluation model trained on EditHF-1M to provide fine-grained, human-aligned scores for image editing across dimensions: visual quality, editing alignment, and attribute preservation.
📥 Model Weights
You can download the pre-trained LoRA checkpoints from the following link: EditHF
📦 Installation
git clone https://github.com/IntMeGroup/EditHF.git
cd EditHF
pip install requirements.txt⚡ Quick Start
python inference.py \
--source_image "/path/to/source.jpg" \ # Path to the original/source image
--edited_image "/path/to/edited.jpg" \ # Path to the edited/target image
--instruction "Editing instruction" \ # Editing instruction describing desired modifications
--peft_dir "lora_checkpoints_visual" \ # Directory containing LoRA checkpoints and MLP head.
--mode visual # Evaluation dimension: 'visual' for visual quality, 'alignment' for editing instruction alignment, 'preservation' for attribute preservationEditHF-Reward is a reward modeling approach that utilizes EditHF signals to improve text-guided image editing models through reinforcement learning.
📥 Model Weights
You can download the advanced image editing model Qwen-Image-Edit refined with our EditHF-Reward from the following link: Qwen-Image-Edit(EditHF-Reward)
⚡ Quick Start
pip install diffusers==0.36.0
python Qweninfer.py \
--source_image "/path/to/source.jpg" \ # Path to the original/source image
--instruction "apply a warm cinematic tone" \ # Editing instruction describing desired modifications
--output "/path/to/output.jpg" \ # Output image path🎨 Editing Examples
If you find our work useful, please cite our paper as:
@article{xu2026edithf1mmillionscalerichhuman,
title={EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing},
author={Zitong Xu and Huiyu Duan and Zhongpeng Ji and Xinyun Zhang and Yutao Liu and Xiongkuo Min and others},
year={2026},
journal={arXiv preprint arXiv:2603.14916},
}