PanoWorld: Panoramic World Generation
Insta360 · hf · 2026-07-13
This work introduces PanoWorld for real-world panoramic generation, aiming to solve the long-range memory problem in panoramic world models.
Key contributions include:
- Leveraging the rotational equivariance of omnidirectional representations to convert rotations into implicit geometric transformations.
- Designing DPRC (Dense Panoramic Ray-Conditioning) and GMA (Geometry-aware Memory Augmentation).
- Proposing a three-stage training pipeline to progressively optimize each module.
- Constructing the World360 dataset, featuring real drone panoramic videos and high-quality simulated clips generated via AirSim360.
Extensive experiments on World360 show that PanoWorld significantly outperforms existing methods. The models, training code, and datasets are planned for public release.
More from Multimodal
- Dense instance segmentation via Muse Wonder + SAM 3.1 on Meta's Model API — nikhilaravi · 2026-10-26
- LichtFeld Studio v0.5.4.3 ships community-built SteadySplats, POPSpa and Voxelization — janusch_patas · 2026-10-26
- One prompt to a 2-3 min narrated short: hands-on test of Antigravity video generation — AI_Andrew · 2026-10-26
- Creator makes 10,564-frame animated short entirely with Gemini toolchain — AI_Andrew · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26