STEP 1 / THE IDEA
It builds the map while the camera is still moving
Most 3D reconstruction workflows collect everything first and process it later. LingBot-Map is designed for streaming reconstruction: each new RGB frame updates its estimate of the camera pose and the surrounding geometry.
Robbyant describes it as a feed-forward 3D foundation model built around a Geometric Context Transformer. The reported target is roughly 20 FPS at 518 × 378 resolution, with support for sequences longer than 10,000 frames.
VIDEO FRAMES → CAMERA POSE + DEPTH + POINT CLOUD → LIVE 3D VIEWSTEP 2 / WHY IT HOLDS TOGETHER
Three kinds of memory keep the scene grounded
The model separates spatial context into three jobs instead of asking one giant attention window to remember everything.
ANCHOR CONTEXT
Keep one stable coordinate reference
The anchor gives later frames a shared origin so the reconstruction does not constantly redefine where the scene begins.
POSE-REFERENCE WINDOW
Track the geometry near the current frame
A local reference window keeps nearby camera movement and dense scene details connected.
TRAJECTORY MEMORY
Correct long-range drift
A compact history of the path helps the model stay coherent as the sequence grows.
STEP 3 / BEFORE YOU INSTALL
Know what kind of tool this is
LingBot-Map is an open research implementation for a local Python and CUDA environment. It is not a phone app, a hosted photogrammetry service, or a replacement for a licensed survey workflow.
Use the official repository as the final authority for supported CUDA, PyTorch, checkpoint, and memory requirements. The release changes quickly, so copy commands from the current README when your local setup differs.
- A CUDA-capable NVIDIA GPU and a compatible local development environment
- Python 3.10 in the official setup path
- The model checkpoint from Robbyant’s Hugging Face repository
- A short, steady image sequence or video for the first test
STEP 4 / OFFICIAL SETUP PATH
Install the model and launch the example viewer
The official README recommends a conda environment, PyTorch with CUDA, the editable project install, and FlashInfer for efficient streaming attention.
git clone https://github.com/Robbyant/lingbot-map.git
cd lingbot-map
conda create -n lingbot-map python=3.10 -y
conda activate lingbot-map
pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
pip install -e .
pip install --index-url https://pypi.org/simple flashinfer-pythonImportant: Match the PyTorch command to your CUDA environment. The repository documents an SDPA fallback when FlashInfer is unavailable.
STEP 5 / FIRST RUN
Start with the included courthouse scene
Download the balanced or long-sequence checkpoint from the official model page, replace the placeholder path below, and run the included example. The viewer opens locally in your browser.
- Use sky masking for outdoor scenes to reduce stray sky points.
- For very long clips, use the repository’s windowed inference mode.
- Begin with a short test before committing a full drone sequence.
python demo.py --model_path /path/to/lingbot-map.pt \
--image_folder example/courthouse --mask_skySTEP 6 / DORIAN’S TEST
Try it as an R&D layer—not a client deliverable
Capture a slow, stable orbit or walkthrough with strong overlap and limited motion blur. Extract frames, run the sequence, then compare the estimated path and point cloud with the source footage.
The useful question is not “does this replace photogrammetry?” It is “can this give me fast spatial context while I am still capturing?” Treat accuracy, scale, and repeatability as things to validate—not assumptions.
CAPTURE → RECONSTRUCT → INSPECT DRIFT → ADJUST PATH → RETESTOfficial Resources
KEEP LEARNING
One guide down. The full library is waiting.
Go back to the Dorian Drone resource library for the Part 107 roadmap, creator systems, drone workflows, and new field guides.
Browse All Free GuidesNo paywall. No fake results. No fluff.