Reproduce “WiseEdit: Benchmarking Cognition- and Creativity-Informed Image Editing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kaihang Pan, Weile Chen, Haiyi Qiu, Qifan Yu, Wendong Bu, Zehan Wang, Yun Zhu, Juncheng Li and 1 more
Researchers from Zhejiang University and Shanghai AI Lab introduced WiseEdit, a benchmark that comprehensively evaluates image editing models across cognitive steps like awareness, interpretation, and imagination, utilizing diverse human knowledge types. Evaluation of 21 models on this benchmark demonstrated a persistent performance gap between open and closed-source systems and highlighted limitations in knowledge fidelity and compositional creativity.
Key insights
- Models integrating visual comprehension (VLMs) with diffusion models consistently outperform purely diffusion-based approaches, highlighting the criticality of robust visual understanding for advanced image editing.
- A significant performance gap persists between closed-source models, which substantially outperform open-source counterparts, indicating their continued leadership in advanced AI capabilities.
- Knowledge-based reasoning and compositional creativity remain key bottlenecks, as models struggle with fully applying world knowledge to follow complex instructions and preserving subject identity during creative transformations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.