Reproduce “BFS: Back-to-Front Layered Image Synthesis via Knowledge Transfer” 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
Kyoungkook Kang, Gyujin Sim, Sunghyun Cho
The BFS framework from Samsung and POSTECH introduces a dual-branch diffusion architecture with a two-stage semi-supervised learning strategy for back-to-front layered image synthesis. It generates high-quality, editable RGBA foreground layers, including associated visual effects, that seamlessly harmonize with a given background, outperforming existing methods in aesthetic quality and human preference while addressing data scarcity challenges.
Key insights
- Knowledge transfer from relatively abundant unlayered image data can effectively guide the learning of complex layered image synthesis, alleviating issues from layered data scarcity.
- Enforcing compositional consistency through a dedicated loss between simultaneously generated foreground and composite branches is critical for producing harmonious and realistic layered images.
- A two-stage semi-supervised learning approach, combining synthetic layered data for structure and unlayered data for realism, provides a robust pathway to high-quality layered generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.