---
library_name: peft
pipeline_tag: image-text-to-text
---

# EditHF

EditHF is an MLLM-based evaluation model introduced in the paper [EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing](https://huggingface.co/papers/2603.14916). 

It is designed to provide fine-grained, human-aligned scores for text-guided image editing across three dimensions: **visual quality**, **editing alignment**, and **attribute preservation**. The model was trained on the **EditHF-1M** dataset, which contains over 29M human preference pairs.

## Resources
- **Paper:** [EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing](https://huggingface.co/papers/2603.14916)
- **GitHub Repository:** [IntMeGroup/EditHF](https://github.com/IntMeGroup/EditHF)

## Sample Usage

To use EditHF for evaluating image editing results, you can use the inference script provided in the official repository:

```bash
python inference.py \
    --source_image "/path/to/source.jpg" \
    --edited_image "/path/to/edited.jpg" \
    --instruction "Editing instruction" \
    --peft_dir "lora_checkpoints_visual" \
    --mode visual
```

The `--mode` parameter can be set to:
- `visual`: Evaluates visual quality.
- `alignment`: Evaluates alignment with the editing instruction.
- `preservation`: Evaluates the preservation of source image attributes.

## Citation
```bibtex
@article{edithf1m,
  title={EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing},
  author={...},
  journal={arXiv preprint arXiv:2603.14916},
  year={2026}
}
```