Instructions to use shuinb/Quantized-LingBot-Map with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LingBot-Map
How to use shuinb/Quantized-LingBot-Map with LingBot-Map:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Quantized LingBot-Map
ไธญๆ ยท GitHub code and full instructions
This is the canonical model repository. The previous Dataset repository is retained for existing links.
The calibrated DSFQ model for streaming 3D mapping, with source code and recorded results. DSFQ combines Hadamard rotation, channel smoothing and learned weight/activation clipping for W4A4 GCT Linear layers.
Files
| Path | Content |
|---|---|
| models/dsfq/model.pt | Standalone packed DSFQ checkpoint, about 1.91 GB |
| models/dsfq/model_info.json | Precision, calibration settings and upstream attribution |
| source/Quantized-LingBot-Map.zip | Clean source project with English and Chinese READMEs |
| results/xyz.json | TUM XYZ, 732-frame ground-truth comparison |
| results/indoor.json | Indoor Travel, 5,000-frame agreement with BF16 |
| results/calibration_audit.json | Direct W4A4 calibration and KV-precision diagnostic |
assets/, configs/ |
Visualizations, icon and calibration/streaming configurations |
The published checkpoint quantizes 192 GCT Linear layers. DINO stays BF16; camera and depth heads stay FP32. The default quantized runtime uses FP4 KV storage and BF16 SDPA. Calibration uses 320 independent TUM desk/rpy frames, six streaming states per block and 15 reconstruction epochs. XYZ and Indoor are development/evaluation sequences, not untouched blind tests.
Download and use
Install the project following its GitHub README, then run from the project directory:
hf download shuinb/Quantized-LingBot-Map models/dsfq/model.pt \
--local-dir .
export CUDA_VISIBLE_DEVICES=0
python -m qmap infer --method dsfq --deployment models/dsfq/model.pt \
--images data/xyz/images --start 64 --frames 732 --output runs/xyz_dsfq
For your own chronologically named images, change --images, use --start 0 and set --frames to the image count. Each inference process uses one visible GPU. The automatic backend uses native S4 on SM8x GPUs such as RTX 3080, and 4-bit codes with INT8 GEMM on other supported CUDA GPUs. Native S4 requires the CUDA toolkit and CUTLASS; see GitHub for setup.
The BF16 reference uses the original robbyant/lingbot-map checkpoint. Direct calibrated W4A4 can be reproduced with python -m qmap calibrate --method w4a4 using the source project's independent calibration configuration.
Reconstruction gallery

Generated with this checkpoint: GPU0 ran 732 XYZ frames and GPU1 ran 496 Indoor frames, with 294ร518 inputs. The curated maps show XYZ frames 200โ327 and Indoor frames 1216โ1279; each uses 64 sampled views, pixel stride 2, and removes the lowest 20% of depth confidence per frame. The GIFs orbit a fixed predicted point map and do not represent online reconstruction speed. RGB/depth panels are unmasked, with depth colors saturated at the 2nd/98th percentiles. No GT geometry, mesh filling or recalibration is used. Rendering records ยท Reproduction commands.
After inference, generate your own panel, PLY and optional GIF on CPU:
CUDA_VISIBLE_DEVICES='' python -m qmap gallery --run runs/xyz_dsfq \
--output runs/xyz_gallery --title 'TUM XYZ / workspace' \
--frame 200 --map-start 136 --map-end 264 --pixel-stride 2 \
--yaw -25 --elevation 5 --animation-frames 40
Frame arguments are zero-based indices within the run. The image resolution is a display choice; inference remains 294ร518.
Recorded results
TUM XYZ, lower is better:
| Method | GT ATE (cm) | Translation RPE (cm) | Depth AbsRel (%) |
|---|---|---|---|
| BF16 | 3.312 | 0.826 | 4.149 |
| Direct LWC/LAC W4A4 + FP4 KV | 8.805 | 3.844 | 5.600 |
| DSFQ + FP4 KV | 3.389 | 1.561 | 4.206 |
XYZ frames 164, 429 and 694: each method/frame is divided by its own median depth and all methods share the BF16 2nd/98th-percentile color limits within each row, without confidence masking. This shows relative shape, not absolute scale or GT error. PDF.
Indoor has no ground truth; its metrics measure agreement with BF16. DSFQ measured a 5.876 GB sampled process peak and 11.987 FPS minimum / 12.477 FPS median across three 5,000-frame runs on RTX 5090. This GPU uses the INT8 compatibility backend. An RTX 3080 native S4 test covered 248 frames at 5.769 GB and 5.611 FPS. Full runtime differences and measurement scopes are recorded in the source README and result files.
License and attribution
The source and derived checkpoint retain the upstream Apache-2.0 terms; see LICENSE and NOTICE. The checkpoint is derived from LingBot-Map; quantization follows QuantVGGT. Visualizations use the attributed TUM RGB-D and Indoor Travel evaluation sequences. Original datasets retain their own terms and are obtained separately.
Model tree for shuinb/Quantized-LingBot-Map
Base model
robbyant/lingbot-map



