```bash
git clone git@github.com:OpenSenseNova/SenseNova-SI.git
cd SenseNova-SI/
uv sync --extra cu124 # or one of [cu118|cu121|cu124|cu126|cu128|cu129], depending on your CUDA version
source .venv/bin/activate
```
#### Hello World
A simple image-free test to verify environment setup and download the model.
```bash
python example.py \
--question "Hello" \
--model_path sensenova/SenseNova-SI-1.4-InternVL3-8B
```
#### Switching Between Supported Models
We fully support multiple model architectures.
To use a different model, simply change the value of the --model_path argument, no other code changes are required.
To use BAGEL-MoT:
```bash
--model_path sensenova/SenseNova-SI-1.1-BAGEL-7B-MoT
```
To use Qwen3-VL:
```bash
--model_path sensenova/SenseNova-SI-1.3-Qwen3-VL-8B
```
### Examples
For more examples, see [example](docs/en/example.md).
#### Example for BAGEL generation
To run the image generation example specifically for the BAGEL-7B-MoT structure, use the following command:
```bash
python example_bagel.py \
--model_path sensenova/SenseNova-SI-1.1-BAGEL-7B-MoT \
--prompt "A chubby cat made of 3D point clouds, stretching its body, translucent with a soft glow." \
--mode generate
```
Use `--mode think_generate` to activate the thinking before generation. Below is a comparison of two modes for the same prompt：
| mode=generate |
mode=think_generate |
|
|
#### Example 1
This example is from [SITE-Bench](https://github.com/wenqi-wang20/SITE-Bench):
```bash
python example.py \
--image_paths examples/Q1_1.png \
--question "Question: Consider the real-world 3D locations of the objects. Which is closer to the sink, the toilet paper or the towel?\nOptions: \nA. toilet paper\nB. towel\nGive me the answer letter directly. The best answer is:" \
--model_path sensenova/SenseNova-SI-1.5-InternVL3-8B
# --model_path sensenova/SenseNova-SI-1.3-Qwen3-VL-8B
```
Details of Example 1
Q: Question: Consider the real-world 3D locations of the objects. Which is closer to the sink, the toilet paper or the towel?\nOptions: \nA. toilet paper\nB. towel\nGive me the answer letter directly. The best answer is:
GT: A
#### Example 2
This example is from [MMSI-Bench](https://github.com/InternRobotics/MMSI-Bench):
```bash
python example.py \
--image_paths examples/Q2_1.png examples/Q2_2.png \
--question "If the landscape painting is on the east side of the bedroom, where is the window located in the bedroom?\nOptions: A. North side, B. South side, C. West side, D. East side\nAnswer with the option's letter from the given choices directly. Enclose the option's letter within ``." \
--model_path sensenova/SenseNova-SI-1.5-InternVL3-8B
# --model_path sensenova/SenseNova-SI-1.3-Qwen3-VL-8B
```
Details of Example 2
Q: If the landscape painting is on the east side of the bedroom, where is the window located in the bedroom?\nOptions: A. North side, B. South side, C. West side, D. East side\nAnswer with the option's letter from the given choices directly. Enclose the option's letter within ``.
GT: C
#### Example 3
This example is from [MMSI-Bench](https://github.com/InternRobotics/MMSI-Bench) and test the model's capability in open-ended short-answer questions:
```bash
python example.py \
--image_paths examples/Q3_1.png examples/Q3_2.png examples/Q3_3.png \
--question "The robot is making tea. What is the order in which the pictures were taken?" \
--model_path sensenova/SenseNova-SI-1.3-InternVL3-8B
```
Details of Example 3
Q: The robot is making tea. What is the order in which the pictures were taken?
GT: Second, first, third
#### Example 4
This example demonstrates the model's **grounding** capability, from [RefCOCO](https://github.com/lichengunc/refer):
```bash
python example.py \
--image_paths examples/Q4.png \
--question "Please provide the bounding box coordinate of the region this sentence describes: [blue shirt lady]" \
--model_path sensenova/SenseNova-SI-1.4-InternVL3-8B
```
Details of Example 4
Q: Please provide the bounding box coordinate of the region this sentence describes: <ref>blue shirt lady</ref>
GT: [0.096234, 0.161229, 0.436516, 1.000000]
#### Example 5
This example demonstrates the model's **depth estimation** capability:
```bash
python example.py \
--image_paths examples/Q5.png \
--question "Identify the minimal distance between the point and the camera, in meters." \
--model_path sensenova/SenseNova-SI-1.4-InternVL3-8B
```
Details of Example 5
Q: Identify the minimal distance between the point and the camera, in meters.
GT: 4.4
#### Example 6
This example demonstrates the model's capability in **solid geometry(Three views)**:
```bash
python example.py \
--image_paths examples/Q6.png \
--question "Enclose your thinking process in tags and your final answer in " \
--model_path sensenova/SenseNova-SI-1.5-InternVL3-8B
```
Details of Example 6
Q: Enclose your thinking process in <think> </think> tags and your final answer in <answer> </answer>
GT: D
#### Example 7
This example demonstrates the model's capability in **solid geometry(Nets of 3D Shapes)**:
```bash
python example.py \
--image_paths examples/Q7.png \
--question "请将你的思考过程放在标签内，并将你的最终答案放在标签内。" \
--model_path sensenova/SenseNova-SI-1.5-InternVL3-8B
```
Details of Example 7
Q: Enclose your thinking process in <think> </think> tags and your final answer in <answer> </answer>
GT: D
#### Test Multiple Questions in a Single Run
Prepare a file similar to [examples/examples.jsonl](examples/examples.jsonl), where each line represents a single question.
The model is loaded once and processes questions sequentially. The questions remain independent of each other.
> For more details on the `jsonl` format, refer to the documentation for [Single-Image Data](https://internvl.readthedocs.io/en/latest/get_started/chat_data_format.html#single-image-data) and [Multi-Image Data](https://internvl.readthedocs.io/en/latest/get_started/chat_data_format.html#multi-image-data).
```bash
python example.py \
--jsonl_path examples/examples.jsonl \
--model_path sensenova/SenseNova-SI-1.3-InternVL3-8B
```
### Training
#### 1. Download Dataset
User may choose to download [SenseNova-SI-800K](https://huggingface.co/datasets/sensenova/SenseNova-SI-800K) (a downsampled subset of SenseNova-SI-8M, specifically designed for studying scaling laws) or [SenseNova-SI-8M](https://huggingface.co/datasets/sensenova/SenseNova-SI-8M) (official full-scale training dataset).
Download [SenseNova-SI-800K](https://huggingface.co/datasets/sensenova/SenseNova-SI-800K) into `training/data/`:
```bash
pip install huggingface_hub
huggingface-cli download sensenova/SenseNova-SI-800K --repo-type dataset --local-dir training/data/SenseNova-SI-800K
```
Download [SenseNova-SI-8M](https://huggingface.co/datasets/sensenova/SenseNova-SI-8M) into `training/data/`:
```bash
pip install huggingface_hub
huggingface-cli download sensenova/SenseNova-SI-8M --repo-type dataset --local-dir training/data/SenseNova-SI-8M
```
#### 2(a). Training with InternVL
**Download pretrained model**
Download [InternVL3-8B](https://huggingface.co/OpenGVLab/InternVL3-8B) into `training/pretrained_models/`:
```bash
huggingface-cli download OpenGVLab/InternVL3-8B --local-dir training/pretrained_models/OpenGVLab/InternVL3-8B
```
**Install dependencies**
```bash
conda create -n internvl python=3.10 -y
conda activate internvl
pip install uv
uv pip install -r training/intern_vl/requirements.txt
uv pip install flash-attn==2.3.6
```
**Run training**
```bash
bash training/intern_vl/internvl_chat/shell/sensenova_si_800K_internvl3_8b.sh #Train with SenseNova-SI-800K data
bash training/intern_vl/internvl_chat/shell/sensenova_si_8M_internvl3_8b.sh # Or train with SenseNova-SI-8M data
```
#### 2(b). Training with Qwen3-VL
The training framework is [lmms-engine](https://github.com/EvolvingLMMs-Lab/lmms-engine), included as a git submodule under `training/lmms-engine/`.
**Download pretrained model**
Download [Qwen3VL-8B](https://github.com/QwenLM/Qwen3-VL) into `training/pretrained_models/`:
```bash
huggingface-cli download Qwen/Qwen3-VL-8B-Instruct --local-dir training/pretrained_models/Qwen/Qwen3-VL-8B-Instruct
```
**Install dependencies**
```bash
# Initialize the lmms-engine submodule (first time only)
git submodule update --init --recursive
conda create -n qwen3vl python=3.10 -y
uv pip install -e training/lmms-engine
# Optional: Performance optimizations
uv pip install flash-attn --no-build-isolation
uv pip install liger-kernel
```
**Preprocess dataset**
Convert `SenseNova-SI-800K.jsonl` and `SenseNova-SI-8M.jsonl` to Qwen3-VL training format:
```bash
python training/qwen3_vl/preprocess_sensenova_si_dataset.py \
--src data/SenseNova-SI-800K/SenseNova-SI-800K.jsonl \
--dst data/SenseNova-SI-800K/SenseNova-SI-800K_qwen3vl_format.jsonl #Preprocess SenseNova-SI-800K data
python training/qwen3_vl/preprocess_sensenova_si_dataset.py \
--src data/SenseNova-SI-8M/SenseNova-SI-8M.jsonl \
--dst data/SenseNova-SI-8M/SenseNova-SI-8M_qwen3vl_format.jsonl #Preprocess SenseNova-SI-8M data
```
**Prepare dataset YAML**
see [training/qwen3_vl/data_800K.yaml](training/qwen3_vl/data_800K.yaml) and [training/qwen3_vl/data_8M.yaml](training/qwen3_vl/data_8M.yaml)
**Configure training**
See [training/qwen3_vl/train_config_800K.yaml](training/qwen3_vl/train_config_800K.yaml) and [training/qwen3_vl/train_config_8M.yaml](training/qwen3_vl/train_config_8M.yaml)
**Run training**
```bash
# Single node, 8 GPUs (default)
bash training/qwen3_vl/run.sh 800K #Train with SenseNova-SI-800K data
bash training/qwen3_vl/run.sh 8M # Or train with SenseNova-SI-8M data
```
#### 2(c). Training with Bagel
**Download pretrained model**
Download [BAGEL-7B-MoT](https://huggingface.co/ByteDance-Seed/BAGEL-7B-MoT) into `training/pretrained_models/`:
```bash
huggingface-cli download ByteDance-Seed/BAGEL-7B-MoT --local-dir training/pretrained_models/BAGEL-7B-MoT
```
**Install dependencies**
```bash
conda create -n bagel python=3.10 -y
conda activate bagel
pip install uv
uv pip install -r training/bagel/requirements.txt
uv pip install flash_attn==2.5.8 --no-build-isolation
```
**Run training**
```bash
bash training/bagel/scripts/train_sensenova_si_800K.sh #Train with SenseNova-SI-800K data
bash training/bagel/scripts/train_sensenova_si_8M.sh # Or train with SenseNova-SI-8M data
```
For details on training hyperparameters (learning rate, batch size, FSDP config, etc.), refer to [training/bagel/TRAIN.md](training/bagel/TRAIN.md).
### Evaluation
To reproduce the benchmark results above, please refer to [EASI](https://github.com/EvolvingLMMs-Lab/EASI) to evaluate SenseNova-SI on mainstream spatial intelligence benchmarks.
EASI supports over 20 spatial intelligence models and more than 20 spatial benchmarks, offering Docker for one-click spatial intelligence evaluation.
## Acknowledgements
This project includes code that is modified from the original code by the BAGEL, InternVL, lmms-engine team.
* Source repository: [BAGEL](https://github.com/bytedance-seed/BAGEL), [InternVL](https://github.com/opengvlab/internvl), [lmms-engine](https://github.com/EvolvingLMMs-Lab/lmms-engine)
We gratefully acknowledge the authors and contributors for their work.
Please refer to the original repositories for full details, updates, and licensing information.
## 🖊️ Citation
```bib
@InProceedings{sensenova-si,
title = {Scaling Spatial Intelligence with Multimodal Foundation Models},
author = {Cai, Zhongang and Wang, Ruisi and Gu, Chenyang and Pu, Fanyi and Xu, Junxiang and Wang, Yubo and Yin, Wanqi and Yang, Zhitao and Wei, Chen and Sun, Qingping and Zhou, Tongxi and Li, Jiaqi and Pang, Hui En and Qian, Oscar and Wei, Yukun and Lin, Zhiqian and Shi, Xuanke and Deng, Kewang and Han, Xiaoyang and Chen, Zukai and Fan, Xiangyu and Deng, Hanming and Lu, Lewei and Pan, Liang and Li, Bo and Liu, Ziwei and Wang, Quan and Lin, Dahua and Yang, Lei},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2026}
}
```
