T3lescope: Arbitrary-Resolution High-Fidelity Generative Surface Reconstruction from Images
AuthorsAtsuhiro Noguchi, Tianhan Xu, Yiming Liang, Yuta Kikuchi, Masahiro Ishiyama, Shintaro Takagi, Hitoshi Murai, Eiichi Matsumoto
AffiliationsPreferred Networks, Inc · Cascade level · The University of Tokyo · Oct 2026 Upscale Upscale Cascade level · Generative reconstruction 185 views 5187 × 3361 resolution Reconstructed mesh (three cascade levels) Figure 1: High-fidelity mesh reconstruction without per-scene optimization. Our method takes multi
T3lescope uses a scalable coarse-to-fine generative process to reconstruct detailed, globally consistent 3D surfaces from posed images across scenes of vastly different sizes.
Key results
Finest cascade level with 32 input views.
Mean distance in meters at the finest cascade level.
Mean score at the finest cascade level.
Approximate speedup using a 2-by-2 rather than 6-by-6 token attention window.
Minutes per scene on one NVIDIA H200.
What the paper found
T3lescope reconstructs detailed 3D meshes from posed images without per-scene optimization by reusing one fixed-resolution generator in a coarse-to-fine cascade: each finer level halves cell size, inherits the parent geometry, and denoises it using image evidence at a matching scale. Built on TRELLIS.2, it combines DINOv3 features with voxel-queried attention and cached ConvNeXt features; Depth Anything 3 helps place the coarsest cells but is not fused into the output mesh. On ScanNet++, the finest level reaches an F@2 score of 0.7155 with 32 input views. On city blocks, it achieves a 0.645 m Chamfer distance and 0.367 F@10, outperforming the tested baselines, whose Chamfer distances are about 5 m or higher. Using a two-by-two token attention window makes mesh-stage sampling roughly 4× faster than the tested 6-by-6 window, with comparable reconstruction quality. The trade-off is compute: reconstruction through the finest level takes 89 minutes per scene on an NVIDIA H200. The method also reconstructs challenging glossy and transparent surfaces, though distant objects in unbounded scenes remain a limitation.
Original abstract
We reconstruct high-fidelity 3D scene meshes from posed multi-view images without per-scene optimization, across scales ranging from single objects to large outdoor scenes. Per-scene optimization methods lack the learned 3D prior needed when observations are sparse or surfaces are glossy or transparent. Existing generative methods leverage such priors to complete geometry in sparsely observed regions, but typically operate at a fixed resolution over a limited spatial extent, trading spatial coverage against detail. Reconstructing a large scene therefore often requires partitioning it into independently processed overlapping local regions, making it difficult to maintain global geometric consistency. To address these issues, we propose T3lescope, which applies 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, so no hierarchy is fixed during training, and the number of levels, cell scales, and cell locations are determined at inference time. On indoor, outdoor, and city-scale scenes, T3lescope outperforms feed-forward and generative baselines, matches or surpasses per-scene optimization, and recovers fine structures as well as glossy and transparent surfaces. These results show that our method generalizes across diverse scenes, view counts, and image resolutions. Project page: https://pfnet-research.github.io/t3lescope/
Read the original paperMore in Generative Models
Browse all 72 papers →Building Rome from a Single Image
Jiraphon Yenphraphai, Fang Li, Tianshuo Xu, Depu Meng, Quentin Herau, Yihan Hu, Raymond A. Yeh, Wei Zhan
The method turns a single photograph into a detailed, complete 3D world by extending object-generation priors to large outdoor scenes.
UniSlider: Perceptually Uniform Sliders for Continuous Image Editing
David Serrano-Lozano, Duygu Ceylan, Yannick Hold-Geoffroy, Iliyan Georgiev, Javier Vazquez-Corral, Anna Frühstück
UniSlider makes image-editing sliders feel more predictable by ensuring that each movement produces a roughly equal perceptual change.
Empirical Variational Autoencoder
Kaede Shiohara
EVA makes VAEs generate high-quality images and sounds faster by replacing their fixed Gaussian latent prior with a learned, self-predictive one.