Reproduce “ProEdit: Inversion-based Editing From Prompts Done Right” 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
Zhi Ouyang, Dian Zheng, Xiao-Ming Wu, Jian-Jian Jiang, Kun-Yu Lin, Jingke Meng, Wei-Shi Zheng
ProEdit introduces a training-free approach to enhance inversion-based visual editing in flow-based generative models by precisely managing source image information injection. It achieves superior fidelity and control for text-driven attribute changes in images and videos, showing quantitative improvements in metrics like CLIP similarity and background preservation.
Key insights
- Excessive source information injection in both latent distribution and attention mechanisms is identified as the primary cause of editing failures in current inversion-based methods.
- Decoupling the preservation of non-edited content from the modification of edited content is crucial for achieving high-fidelity, controllable visual edits.
- Targeted perturbation of the latent distribution in edited regions and selective mixing of attention features based on a mask can effectively mitigate the over-reliance on source information without additional training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.