Reproduce “TanDiT: Tangent-Plane Diffusion Transformer for High-Quality 360° Panorama Generation” 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
Hakan Çapuk, Andrew Bond, Muhammed Burak Kızıl, Emir Göçen, Erkut Erdem, Aykut Erdem
TanDiT generates high-quality 360° panoramas by decomposing scenes into structured tangent-plane views for generation by a Diffusion Transformer, followed by an equirectangular-conditioned refinement step to ensure global coherence. The method demonstrates improved consistency across the entire panoramic field of view and achieves strong user preference over prior approaches.
Key insights
- Operating on undistorted tangent-plane views, arranged in a structured grid, allows Diffusion Transformers to effectively generate high-quality panoramic content by leveraging their native spatial modeling capabilities.
- A dedicated equirectangular-conditioned refinement stage is essential to eliminate stitching artifacts and ensure global coherence, especially when combining individually generated tangent planes.
- Standard image quality metrics are insufficient for 360° panoramas, necessitating specialized evaluation metrics like TangentIS and TangentFID that account for consistency across all regions, including poles.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.