T3lescope: Arbitrary-Resolution High-Fidelity Generative Surface Reconstruction from Images
T3lescope reconstructs high-fidelity 3D meshes from multi-view images at arbitrary scale without per-scene optimization.
T3lescope reconstructs high-fidelity 3D scene meshes from posed multi-view images without per-scene optimization, from single objects to city-scale scenes. A single fixed-resolution generator runs in an inference-time coarse-to-fine cascade: a coarse level sets layout, and finer levels denoise inherited geometry inside smaller cells. Weights are shared across scales and trained on individual cells, so hierarchy is chosen only at inference. The authors report it outperforms feed-forward and generative baselines and matches or surpasses per-scene optimization, including glossy and transparent surfaces.
- Feed-forward coarse-to-fine cascade uses one shared generator at every scale.
- Inference chooses level count, cell scale, and cell locations; training fixes no hierarchy.
- Matches or beats per-scene optimization on indoor, outdoor, and city scenes.
- Recovers fine structure plus glossy and transparent surfaces from sparse views.
- Organizations
- Preferred Networks
- AI models
- T3lescope
From the original article
1 min read at huggingface.co
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.
By Atsuhiro Noguchi, Tianhan Xu, Yiming Liang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.