Reproduce “CannyEdit: Selective Canny Control and Dual-Prompt Guidance for Training-Free 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
Weiyan Xie, Han Gao, Didan Deng, Kaican Li, April Hua Liu, Yongxiang Huang, Nevin L. Zhang
CannyEdit combines selective Canny edge guidance outside edit regions with local and global prompts for training-free, region-based image editing. It uses source contours to preserve context while leaving requested changes freer and prompt guidance to connect them. For adding two objects on RICE-Bench-Add2, it reports CF 83.76, TA 25.34, and PR 92.56; these results cover that setting, not general editing.
Key insights
- Selectively applying structural control only to unedited regions effectively mitigates boundary artifacts and ensures context fidelity without sacrificing editability in the masked areas.
- A dual-prompt strategy, combining localized editing prompts with a global coherence prompt via an adaptable attention mechanism, is critical for achieving both precise content generation and seamless integration into the broader scene.
- Integrating an automatic mask refinement procedure during the denoising process significantly enhances the usability of training-free editing by allowing flexible input cues like single-point hints and improving object addition accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.