LingBot-Map is a feed-forward 3D foundation model designed for reconstructing scenes from streaming data, offering high-efficiency streaming inference and state-of-the-art reconstruction performance.
Source: Official README and linked documentation View on GitHub →LingBot-Map is being discussed because its innovative Geometric Context Transformer architecture, which unifies coordinate grounding, dense geometric cues, and long-range drift correction. Its high-efficiency streaming inference and superior performance on diverse benchmarks make it a solution for 3D reconstruction tasks.
Source: Official README and linked documentationArchitecturally unifies coordinate grounding, dense geometric cues, and long-range drift correction within a single streaming framework, enhancing the accuracy and behavior under varied inputs of 3D reconstruction.
Source: Official README and linked documentationUtilizes a feed-forward architecture with paged KV cache attention, enabling stable inference at ~20 FPS on high-resolution sequences, making it suitable for real-time applications.
Source: Official README and linked documentationDemonstrates superior performance on diverse benchmarks compared to existing methods, indicating its effectiveness in reconstructing complex scenes.
Source: Official README and linked documentationThe architecture is inferred to be modular, with distinct components for geometric context transformation, streaming inference, and reconstruction. The use of a feed-forward architecture and paged KV cache attention suggests a focus on efficiency and scalability.
Source: Official repository tree and dependency filesPillowhuggingface_hubeinopssafetensorsopencv-pythontqdmscipyvisertrimeshmatplotlibonnxruntimerequestsLingBot-Map is suitable for applications requiring real-time 3D reconstruction from streaming data, such as robotics, augmented reality, and autonomous vehicles. It can be used for tasks like scene understanding, object detection, and path planning.
Source: Official README and linked documentationAs of 2026-09-23, the project has no GitHub Release; pyproject.toml reports package version 0.1.0. Reproduction should pin both commit and checkpoint.
Source: github.com/Robbyant/lingbot-map/releasesstatus: docs-only | checked_at: 2026-09-23 | environment: The official README, pyproject.toml, benchmark notes, model-download links, and GitHub Releases page were reviewed. | limitations: No local installation or runtime verification was performed. CUDA, PyTorch, FlashInfer, Kaolin, checkpoints, datasets, inference, the viewer, benchmarks, offline rendering, and hardware-dependent claims such as 20 FPS were not tested. | observed_output: pyproject.toml reports version 0.1.0 and Python >=3.10. There is no GitHub Release. The README records an SDPA KV-cache fix on 2026-06-28 and a benchmark published on 2026-05-25.
Source: github.com/Robbyant/lingbot-map/releasesMeta VGGT accepts one to hundreds of images and predicts cameras, depth, point maps, and tracks in one feed-forward pass. LingBot-Map uses streaming state, KV cache, keyframes, and window resets for longer sequences. VGGT fits short batches and multi-task geometry; LingBot-Map targets continuous video and long-sequence state.
Source: github.com/facebookresearch/vggtLingBot-Map is a open-source project for 3D reconstruction tasks, offering innovative architecture and high performance. It is particularly suitable for applications requiring real-time reconstruction and is worth watching for further development and adoption.
pyproject.toml reports version 0.1.0 and Python >=3.10. There is no GitHub Release. The README records an SDPA KV-cache fix on 2026-06-28 and a benchmark published on 2026-05-25. No local installation or runtime verification was performed.
The review environment was: The official README, pyproject.toml, benchmark notes, model-download links, and GitHub Releases page were reviewed. lingbot-map behavior outside that scope still needs a local check.