Enhancing In-context Panoramic Generation via Geometric-aware Pretraining

The paper introduces Canvas360, a two-stage framework for in-context panoramic generation that combines geometry-aware pretraining with task-specific fine-tuning. It also presents Canvas360Dataset, a collection of 1 million high-quality paired panoramic samples covering style transfer, inpainting, outpainting, and editing. The method uses parallel depth generation, velocity circular padding, and similarity-loss regularization to model panoramic geometry, object distortion, and global coherence. A unified downstream framework is built through token-level concatenation. According to the paper, Canvas360 achieves particularly strong results on the panorama-specific FAED metric and competitive or leading performance across the reported evaluations.