Researchers have introduced T3lescope, a generative method for reconstructing high-fidelity 3D scene meshes from posed multi-view images without per-scene optimization. The work targets a persistent trade-off in 3D reconstruction: methods that optimize a 3D representation for each scene lack a learned 3D prior when observations are sparse or surfaces are glossy or transparent, while existing generative approaches that supply such a prior typically operate at a fixed resolution over a limited spatial extent.
That fixed-resolution constraint forces large scenes to be split into independently processed overlapping local regions, which makes global geometric consistency difficult to maintain. T3lescope addresses this by applying a single fixed-resolution generator across scene scales in an inference-time coarse-to-fine cascade: a coarse level establishes the scene layout, and finer levels perturb and denoise geometry inherited from the coarser level within progressively finer spatial cells to recover surface detail.
The model is trained on individual cells at multiple scales and shares its weights across all levels. As a result, no hierarchy is fixed during training — the number of levels, cell scales, and cell locations are all determined at inference time.
According to the authors, T3lescope outperforms feed-forward and generative baselines on indoor, outdoor, and city-scale scenes, matches or surpasses per-scene optimization, and recovers fine structures as well as glossy and transparent surfaces. The results suggest the method generalizes across diverse scenes, view counts, and image resolutions.
A project page is available at pfnet-research.github.io/t3lescope.