SHERPA: Seam-aware Harmonized ERP Adaptation for Open-Domain 360$^\circ$ Panorama Generation
Researchers have developed SHERPA, a new framework designed to adapt large-scale text-to-image models for generating 360-degree panoramas. Existing models struggle with the unique topology of equirectangular projection (ERP) panoramas, leading to misalignments, especially at the seams and polar regions. SHERPA addresses this by incorporating frequency-selective RoPE, circular encoding, and a dual-path training scheme to enable the generation of both photorealistic and stylized panoramic scenes. AI
IMPACT Enables more accurate and stylized 360-degree panorama generation from text-to-image models.