Canvas360 learns geometry-aware panoramic generation from 1M paired samples

Canvas360, a two-stage framework for in-context panoramic generation, combines geometry-aware pretraining with task-specific fine-tuning to handle the full 360-degree image space. Researchers introduced Canvas360Dataset, a collection of 1 million high-quality paired panoramic samples covering style transfer, inpainting, outpainting, and editing. The dataset addresses a critical gap: prior panoramic generation work has been hampered by the lack of large-scale, task-specific training data.

The framework enhances text-to-panorama generation through parallel depth generation, velocity circular padding, and similarity loss regularization—techniques designed to learn geometry-aware representations and capture the object distortion inherent in 360-degree imagery. A unified architecture handles multiple downstream tasks via token-level concatenation, eliminating the need for separate models per task. In evaluation, Canvas360 achieved particularly strong performance on the panorama-specific FAED metric, with competitive or leading results across quantitative benchmarks. The preprint was posted to HuggingFace Papers on July 10, 2026.