LingBot-Map streams 3D scene reconstruction at 20 FPS

LingBot-Map is an open-source 3D foundation model designed for streaming 3D reconstruction, aiming to rebuild scenes from long image or video sequences without relying on iterative optimization at inference time. The project, published by the Robbyant Team, centers on what it calls a Geometric Context Transformer, which combines coordinate grounding, dense geometric cues, and long-range drift correction in one streaming system.

According to the project page, the model uses a feed-forward architecture with paged KV-cache attention, delivering stable inference at roughly 20 FPS on 518×378 resolution across sequences longer than 10,000 frames. The team says it outperforms both existing streaming methods and optimization-based approaches on multiple benchmarks.

Recent updates show the project is still moving quickly. On 2026-06-28, the team said it fixed an SDPA KV-cache bug, improving long-sequence performance, though it still recommends the FlashInfer backend. Earlier releases added evaluation scripts for KITTI and Oxford Spires, plus a long-video demo covering about 25,000 frames over a 13-minute indoor walkthrough.

The repository includes checkpoints for different use cases:

• 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 as a stage-1 training checkpoint that can be loaded into the VGGT model for bidirectional inference

Setup is built around Python 3.10, PyTorch 2.8.0, and optional FlashInfer support. For long runs, the project supports windowed inference, keyframe intervals to reduce memory use, and an offline rendering pipeline that outputs headless point-cloud flythrough videos. The team also provides sky masking via an ONNX segmentation model and sample scenes including courthouse, university, loop, and oxford.

Recommended reading

Moonshot’s Kimi K3 packs 2.8T parameters

For users pushing beyond the model’s training range, the repository warns that the default method does not reset state automatically, so very long trajectories may require windowed mode to avoid pose collapse.