Preferred Networks' T3lescope reconstructs high-fidelity 3D scenes from multi-view images without per-scene optimization
pfnet · hf · 2026-10-13
Preferred Networks and the University of Tokyo propose T3lescope, which generates high-fidelity 3D scene meshes from posed multi-view images without per-scene optimization, scaling from single objects to city-scale scenes.
- Approach: a single fixed-resolution generative model runs in an inference-time coarse-to-fine cascade — a coarse level sets the layout, finer levels perturb and denoise inherited geometry within progressively smaller spatial cells; level count, cell scales and locations are decided at inference, with weights shared across all levels.
- Motivation: per-scene optimization lacks learned priors for sparse views or glossy/transparent surfaces; existing generative methods trade coverage against detail and require partitioning large scenes.
- Results: outperforms feed-forward and generative baselines on indoor, outdoor and city-scale scenes, matches or beats per-scene optimization, and recovers fine structures as well as glossy and transparent surfaces.
More from Research
- Harvard & NTT Research: cooperating AI agent populations hit an Allee-effect takeoff threshold — Hidenori8Tanaka · 2026-10-13
- RT-Super trains tumor segmentation from radiology reports, lifting Dice from 16 to 30 without masks — maier_ak · 2026-10-13
- Princeton's QUEEN: a 4B model that plays grandmaster-level chess and explains its moves in plain language — jiqizhixin · 2026-10-13
- Oxford shows neural nets can be encrypted inside model weights, surviving 500 SFT steps — zainhas · 2026-10-13
- XPENG's 4B GroundingPI model beats GPT-6 Astra across 34 grounding benchmarks, boosts robot manipulation — XPENG-AI · 2026-10-13
- XPENG's GroundAnything pairs diffusion parallel decoding with precise grounding, 4.51x faster than AR — XPENG-AI · 2026-10-13