lingbot-map-3d-reconstruction — Claude Code skill: 3D from video

What lingbot-map-3d-reconstruction does

The skill wires the LingBot-Map model into your agent. Inference is feed-forward, so there is no iterative optimization step. You get a point cloud with per-point confidence, camera poses and depth maps. Under the hood sits a Geometric Context Transformer, and paged KV cache attention keeps long sequences workable: the description claims around 20 FPS at 518×378 across sequences past 10,000 frames.

Everything runs through demo.py. Point it at a folder of images with --image_folder or at a video with --video_path plus --fps. For outdoor footage add --mask_sky. Long runs have --keyframe_interval to store every Nth frame, and a windowed path via --mode windowed --window_size 64. The demo serves an interactive viewer in the browser on localhost:8080.

Before you start, set up Python 3.10, PyTorch for CUDA 12.8, and download the checkpoint from HuggingFace or ModelScope. FlashInfer for fast attention is a separate install; without it you fall back to --use_sdpa. If FlashInfer will not build, the demo still runs, just slower — worth checking against your CUDA version before feeding it a long video.