Current 3D reconstruction methods either require per-scene optimization, struggle with sparse views or reflective surfaces, or are limited to a fixed spatial resolution. T3lescope introduces an inference-time coarse-to-fine cascade that uses a single fixed-resolution 3D generator across multiple scales. The coarse level captures global scene layout, and finer levels refine geometry to recover fine detail, glossy surfaces, and transparent objects.
The method is evaluated on ScanNet++, Tanks and Temples, and a city-block dataset. It outperforms feed-forward and generative baselines and matches or surpasses per-scene optimization. This is interesting because it generalizes across scene scales, view counts, and image resolutions without requiring per-scene training.