ImgEdit-Judge

ImgEdit-Judge (re-host)

Unmodified re-host of ImgEdit_Judge, the LLM-as-judge model for ImgEdit-Bench image-editing evaluation. Published for users who cannot access the original local checkpoint. The weights here are byte-identical to the authors' release — this repository only mirrors them for accessibility.

What it is

Given a source image, an edited image, and an edit instruction + rubric, the model rates the edit on a 1–5 scale across per-category rubric dimensions (used to compute the ImgEdit-Bench "Add / Adjust / Extract / Replace / Remove / Background / Style / Hybrid / Action" scores). It is a fine-tune of Qwen2.5-VL-7B.

Source & attribution

• Original work: PKU-YuanGroup — ImgEdit: A Unified Image Editing Dataset and Benchmark (arXiv:2505.20275, GitHub). The checkpoint is published inside the dataset repo sysuyy/ImgEdit under the ImgEdit_Judge/ subfolder (there is no standalone model repo from the authors).
• Base model: Qwen/Qwen2.5-VL-7B-Instruct (Apache-2.0).
• License: Apache-2.0, inherited from the ImgEdit release and the Qwen2.5-VL-7B base.
• Changes: none — weights are re-hosted unmodified.

All credit belongs to the original ImgEdit authors (PKU-YuanGroup) and the Qwen team.

Citation

Please cite the original work (author list on the paper page):

@article{imgedit2025,
  title  = {ImgEdit: A Unified Image Editing Dataset and Benchmark},
  journal= {arXiv preprint arXiv:2505.20275},
  year   = {2025},
  url    = {https://arxiv.org/abs/2505.20275}
}