Make Geometry Matter for Spatial Reasoning
Paper • 2603.26639 • Published • 32
How to use from
SGLangUse Docker images
docker run --gpus all \
--shm-size 32g \
-p 30000:30000 \
-v ~/.cache/huggingface:/root/.cache/huggingface \
--env "HF_TOKEN=<secret>" \
--ipc=host \
lmsysorg/sglang:latest \
python3 -m sglang.launch_server \
--model-path "SuhZhang/GeoSR-Model" \
--host 0.0.0.0 \
--port 30000# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "SuhZhang/GeoSR-Model",
"prompt": "Once upon a time,",
"max_tokens": 512,
"temperature": 0.5
}'Quick Links
GeoSR Model Zoo
This repository hosts the released checkpoints for GeoSR: Make Geometry Matter for Spatial Reasoning.
Paper | Project Page | Code
Checkpoints
| Folder | Branch / task | Notes |
|---|---|---|
GeoSR3D-Model/ |
static branch |
GeoSR checkpoint for static spatial reasoning |
GeoSR4D-Model/ |
dynamic branch |
GeoSR checkpoint for dynamic spatial reasoning |
The model files are stored exactly as exported from our training directories.
Download
Install the client first if needed:
pip install -U huggingface_hub
Download only the static checkpoint:
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='SuhZhang/GeoSR-Model', local_dir='data/models', allow_patterns=['GeoSR3D-Model/*'])"
Download only the dynamic checkpoint:
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='SuhZhang/GeoSR-Model', local_dir='data/models', allow_patterns=['GeoSR4D-Model/*'])"
Download the whole repository:
git lfs install
git clone https://huggingface.co/SuhZhang/GeoSR-Model
Usage
In the main GeoSR code repository:
staticbranch evaluation can useMODEL_PATH=./data/models/GeoSR3D-Modeldynamicbranch evaluation can useGEOSR4D_EVAL_MODEL_PATH=../../../data/models/GeoSR4D-Model
Please refer to the main code repository for full training and evaluation instructions:
- Code: https://github.com/SuhZhang/GeoSR
- Project page: https://suhzhang.github.io/GeoSR/
Citation
@misc{zhang2026geosr,
title={Make Geometry Matter for Spatial Reasoning},
author={Shihua Zhang and Qiuhong Shen and Shizun Wang and Tianbo Pan and Xinchao Wang},
year={2026},
eprint={2603.26639},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.26639}
}
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Install from pip and serve model
# Install SGLang from pip: pip install sglang# Start the SGLang server: python3 -m sglang.launch_server \ --model-path "SuhZhang/GeoSR-Model" \ --host 0.0.0.0 \ --port 30000# Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "SuhZhang/GeoSR-Model", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'