ImgEdit

https://github.com/PKU-YuanGroup/ImgEdit/raw/main/assets/radar.png

https://github.com/PKU-YuanGroup/ImgEdit/raw/main/assets/leaderboard.png

ImgEdit-Bench consists of three key components: a basic editing suite that evaluates instruction adherence, editing quality, and detail preservation across a diverse range of tasks; an Understanding-Grounding-Editing (UGE) suite, which increases task complexity through challenging instructions (e.g., spatial reasoning and multi-object targets) and complex scenes such as multi-instance layouts or camouflaged objects; and a multi-turn editing suite, designed to assess content understanding, content memory, and version backtracking.

ImgEdit: A Unified Image Editing Dataset and Benchmark

ImgEdit-E1

ImgEditを用いて作成されたモデル

DatasetとBenchmark

Dataset

ImgEdit-Bench

ImgEdit_Judge ckptを用いる

ImgEdit Pipeline

https://github.com/PKU-YuanGroup/ImgEdit/raw/main/assets/datapipeline.png

現時点(2025/10/23) SetupがWIPのまま

3つの主要コンポーネントで構成されている

Basic: instruction adherence(指示遵守)/editing quality/detail preservation(細部保持)

Understanding-Grounding-Editing(UGE)

Multi-Turn: content understanding/content memory/version backtracking