A new paper from arXiv proposes turning a single 360-degree photo into a fully explorable 3D environment without the multi-step hand-holding most current methods require.
The system, called Pano2World, takes one indoor panoramic image and outputs what researchers call a "3D Gaussian scene" - a compact geometric representation that supports free-viewpoint navigation. It first builds a rough 3D proxy from the source image, renders that proxy from nearby angles to generate guidance frames, then runs a diffusion model that denoises all target views at once rather than one at a time. A module called Latent Feature Adapter pulls geometric information directly from the model's internal hidden states, skipping a lossy decoding step that comparable approaches rely on. On a standard benchmark for panoramic novel-view synthesis, the method outperforms existing alternatives.
The practical gap being closed here is meaningful: current approaches either chain inpainting steps together - letting errors compound with each iteration - or borrow from video generation models, which are optimized for smooth forward trajectories and struggle to cover a scene from multiple directions at once. Pano2World's joint denoising step addresses both problems in a single forward pass. That matters for anyone building virtual tours, real-estate previews, or spatial AI training data who currently has to babysit a brittle pipeline.
This is academic research, so production readiness is a separate question - but the architecture borrows heavily from diffusion and Gaussian splatting work that has moved from paper to product faster than most expected.