• 2 min read
LingBot-Map streams 3D reconstruction at 20 FPS
LingBot-Map is a feed-forward 3D reconstruction model built for long video streams, with claimed state-of-the-art results and support for 10,000+ frames.
Image: Hacker News
LingBot-Map is an open-source feed-forward 3D foundation model designed for streaming 3D reconstruction from long video sequences. The project, published by the Robbyant Team, centers on a Geometric Context Transformer that combines coordinate grounding, dense geometric cues, and long-range drift correction in one streaming framework.
According to the project page, the model runs at about 20 FPS on 518×378 input and can handle sequences longer than 10,000 frames using a paged KV cache attention setup. The team says it outperforms both prior streaming systems and iterative optimization-based methods across multiple benchmarks.
The repository also documents a fast-moving release cycle. Recent updates include a fix for an SDPA KV cache bug on 2026-06-28, evaluation scripts for KITTI and Oxford Spires on 2026-05-25, and a long-video demo released on 2026-04-29 showing a roughly 25,000-frame, 13-minute indoor walkthrough.
Models and setup
The project offers three checkpoints:
Recommended reading
Satellite startups pull $8.1B in half a year
- lingbot-map-long for long sequences and large-scale scenes
- lingbot-map as the balanced checkpoint used in the paper, benchmark, and offline demo
- lingbot-map-stage1, a stage-1 training checkpoint that can be loaded into the VGGT model for bidirectional inference
Recommended installation uses Python 3.10, PyTorch 2.8.0, and CUDA 12.8. The team recommends FlashInfer for best performance, with SDPA available as a fallback.
Long sequences and rendering
For interactive use, demo.py launches a browser-based viser viewer on http://localhost:8080. Example scenes include courthouse, university, loop, and oxford. For longer runs, the code supports keyframe intervals to reduce memory use, and a windowed mode for sequences beyond 3,000 frames.
An offline pipeline in demo_render/batch_demo.py renders headless point-cloud flythrough videos from either image folders or video files. The featured example processes an indoor video at about 25,000 frames using windowed mode, with optional sky masking via an ONNX segmentation model downloaded from Hugging Face on first use.
The code is released under Apache-2.0.
Frontier Editor
Dan is our resident futurist, covering electric mobility, space exploration, and the smart home. He's interested in atoms just as much as bits. Whether it's a new battery chemistry, a reusable rocket, or a protocol that finally makes IoT devices talk to each other, Dan breaks down the engineering that pushes humanity forward.
via Hacker News
