Reproduce “AlbedoEdit: Unified Instance-Level Video Editing with Albedo Guidance” 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
Xilong Zhou, Bao-Huy Nguyen, Zheng Zeng, Jacob Munkberg, Jon Hasselgren, Thomas Leimkühler, Nima Kalantari, Miloš Hašan and 1 more
AlbedoEdit introduces a unified framework for instance-level video editing, encompassing object insertion, removal, and texture editing, by leveraging intrinsic albedo maps as an illumination-invariant conditioning signal. This approach enables the generation of physically plausible light transport effects for edited content and demonstrates robust generalization from synthetic training to real-world videos.
Key insights
- Intrinsic albedo maps serve as an effective and user-friendly mechanism for specifying precise appearance edits, as they are invariant to lighting conditions and free from complex light interactions.
- Training a generative video model with albedo guidance allows it to implicitly learn and synthesize complex, physically plausible light transport effects for edited content.
- A unified framework, conditioned on intrinsic properties like albedo, can bridge the gap between flexible general models and precise task-specific methods, enhancing both control and realism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.