What is it about?
We train a diffusion model that takes a roughly edited image, and makes it a photorealistic masterpiece while being faithful to the original image!
Featured Image
Why is it important?
Prior work relies on "text" to perform an edit, which is difficult to control precisely, and struggles to be faithful to the original image. With our work, we allow the user to edit their photos intuitively (just do a quick edit without worrying about precision), and the model does the heavy lifting and make your edit neat and realistic.
Perspectives
The key insight is that we used video data to train an "image editing" model. Since in videos, objects move naturally, and it can be used to teach the model on how to re-arrange or edit a photo realistically.
Hadi Alzayer
University of Maryland at College Park
Read the Original
This page is a summary of: Magic Fixup: Streamlining Photo Editing by Watching Dynamic Videos, ACM Transactions on Graphics, July 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3750722.
You can read the full text:
Contributors
The following have contributed to this page
