A simple browser UI for the LingBot-Map streaming 3D reconstructor. Drop in a video or a folder of images, hit run, and view the reconstructed point cloud inline (and download the GLB).
- Three input modes: video, uploaded images, or one of the built-in
example/folders - Auto-download of the checkpoint from HuggingFace
(
robbyant/lingbot-map) - Sensible defaults for first run (120 frames cap, sky mask off, SDPA backend so FlashInfer is not required)
- Result rendered in-browser via Gradio's
Model3Dcomponent, and downloadable as a.glb
setup.bat
run_app.batThat creates a .venv (Python 3.10), installs PyTorch 2.8.0 + CUDA 12.8, the
lingbot-map package + [vis] extras, Gradio, and onnxruntime.
To launch with a public share link:
run_app.bat --shareuv venv --python 3.10 .venv
.venv/Scripts/python.exe -m pip install torch==2.8.0 torchvision==0.23.0 \
--index-url https://download.pytorch.org/whl/cu128
.venv/Scripts/python.exe -m pip install -e ".[vis]" gradio onnxruntime
.venv/Scripts/python.exe app.pyThe UI starts at http://localhost:7860.
- First run is slow β the model is loaded into memory and cached for
subsequent runs (changing the checkpoint,
use_sdpa, or camera iters invalidates the cache). - Long videos: cap with the Max frames field, or leave
keyframe_interval = 0to let it auto-pick. >320 frames will trim KV cache by sub-sampling keyframes. - Outdoor scenes: tick Mask sky. The first time, it downloads a small ONNX model from HuggingFace.
- No FlashInfer? No problem. SDPA is on by default. Untick only if you installed FlashInfer separately.
- Checkpoints live under
checkpoints/. You can also paste a full path into the Custom checkpoint path box to use one you already have.
app.pyβ the Gradio appsetup.bat/run_app.bat/run_app.shβ launcherscheckpoints/β auto-created, holds downloaded.ptfiles