library_name: diffusers
pipeline_tag: image-to-image
EditHF-Reward
This repository contains the weights for EditHF-Reward, an advanced image editing model based on Qwen-Image-Edit that has been refined using reinforcement learning with human-aligned feedback.
This model was introduced in the paper EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing.
Model Description
Recent text-guided image editing (TIE) models often suffer from issues such as artifacts, unexpected edits, or unaesthetic content. EditHF-Reward addresses these challenges by utilizing rewards from EditHF, a multimodal large language model (MLLM) trained on the EditHF-1M dataset.
EditHF-1M is a million-scale dataset featuring over 29 million human preference pairs and 148,000 human mean opinion ratings, covering three critical dimensions:
- Visual Quality: Reducing artifacts and improving aesthetic appeal.
- Instruction Alignment: Ensuring edited images accurately reflect the text instructions.
- Attribute Preservation: Maintaining the identity and characteristics of the original image.
By fine-tuning with reinforcement learning (RL) using these signals, this model achieves significantly improved performance and human alignment compared to its base version.
Resources
- Paper: EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing
- GitHub Repository: IntMeGroup/EditHF
Citation
If you find this model useful in your research, please cite the following paper:
@article{edithf1m,
title={EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing},
author={Luo, Junyu and others},
journal={arXiv preprint arXiv:2603.14916},
year={2026}
}