Niantic Spatial's GroundControl uses satellite images to guide multi-view diffusion for street-level novel view synthesis
kwangmoo_yi · x · 2026-10-29
GroundControl (Niantic Spatial + Georgia Tech + Aalto, accepted to WACV 2027) is a reconstruction-free multi-view diffusion model conditioned on satellite imagery for novel view synthesis.
Problem: multi-view/video diffusion models hallucinate unrealistic spatial layouts when extrapolating views beyond the input coverage.
Approach: satellite images and street-level input images are injected via cross-attention, which retains pretrained multi-view diffusion priors and lets the model generate seamless, realistic street-level views from a handful of source views while following the satellite scene layout — with no 3D reconstruction, only posed street images plus satellite data. Paper and code are available.
More from Multimodal
- KeenFace turns a single photo into a MetaHuman face for UE5 character creation — TinfoilTricorn · 2026-10-29
- AI-generated Batman F1 helmet concept image wows with carbon-fiber realism — Delahuntagram · 2026-10-29
- AI-Generated Fake: Korn's 'Freak on a Leash' as a 1973 Spanish TV Performance — Mother-Pear-2392 · 2026-10-29
- Has Flux.2 Klein 9B Been Cracked? Community Weighs the Fight Against AI-Slop Aesthetics — VGDCMario · 2026-10-29
- Full Local AI Music Video: LongCat-Video-Avatar 1.5 on an RTX 5090 with Sync 3 Lip Sync — ConcentrateDeepTrans · 2026-10-29
- Surreal metallic AI flower art with honeycomb textures and blue gradients — ciguleva · 2026-10-29