VST

<div align='center'> <h1>Visual Spatial Tuning</h1>

</div>

We introduce Visual Spatial Tuning (VST), a comprehensive framework designed to cultivate Vision-Language Models (VLMs) with human-like visuospatial abilities—from spatial perception to advanced reasoning.

---

🔥 News

• Support Qwen3VL training code, see assets/train.md for more details.
• Training code has been updated and verified, please see Train, which is very efficient because of data packing.

💡 Key Highlights

• VST-P: 4.1M samples across 19 skills, spanning single images, multi-image scenarios, and videos—boosting spatial perception in VLMs.
• VST-R: 135K curated samples that teach models to reason in space, including step-by-step reasoning and rule-based data for reinforcement learning.
• Progressive Training Pipeline: Start with supervised fine-tuning to build foundational spatial perception, then reinforce spatial reasoning abilities via RL. VST achieves state-of-the-art results on spatial benchmarks (34.8% on MMSI-Bench, 61.2% on VSIBench) without compromising general capabilities.
• Vision-Language-Action Models Enhanced: The VST paradigm significantly strengthens robotic learning.

---

🏷️ Model Card

Model Name	🤗 HuggingFace
VST-3B-SFT	rayruiyang/VST-3B-SFT
VST-3B-RL	rayruiyang/VST-3B-RL
VST-7B-SFT	rayruiyang/VST-7B-SFT
VST-7B-RL	rayruiyang/VST-7B-RL

shellCopy
# download models into checkpoints
python tools/download_hf_model.py --model_list rayruiyang/VST-3B-SFT rayruiyang/VST-3B-RL rayruiyang/VST-7B-SFT rayruiyang/VST-7B-RL --local_dir checkpoints

<details> <summary>Click to see performance 📈 </summary> <h3>📈 Spatial & General Benchmarks</h3> <table> <tr> <th>Models</th><th>CV</th><th>3DSR</th><th>MMSI</th><th>BLINK</th><th>VSI</th><th>MMStar</th><th>MMB</th><th>RealworldQA</th><th>MMMU</th><th>OCRB</th><th>AI2D</th> </tr> <tr> <td>VST-3B-SFT</td><td>84.4</td><td>54.1</td><td>30.2</td><td>59.1</td><td>57.9</td><td>58.0</td><td>80.9</td><td>68.4</td><td>45.2</td><td>83.7</td><td>82.5</td> </tr> <tr> <td>VST-3B-RL</td><td>84.2</td><td>56.5</td><td>31.3</td><td>57.2</td><td>57.7</td><td>58.9</td><td>80.5</td><td>68.5</td><td>49.8</td><td>80.9</td><td>82.4</td> </tr> <tr> <td>VST-7B-SFT</td><td>85.5</td><td>54.6</td><td>32.0</td><td>62.1</td><td>60.6</td><td>63.1</td><td>83.3</td><td>72.2</td><td>50.6</td><td>85.5</td><td>84.9</td> </tr> <tr> <td>VST-7B-RL</td><td>86.5</td><td>60.1</td><td>34.8</td><td>62.6</td><td>61.2</td><td>63.5</td><td>83.0</td><td>68.5</td><td>49.4</td><td>86.1</td><td>83.5</td> </tr> </table> <h3>📈 VSIBench</h3> <table> <tr> <th>Methods</th><th>Avg.</th><th>Obj. Count</th><th>Abs. Dist.</th><th>Obj. Size</th><th>Room Size</th><th>Rel. Dist</th><th>Rel. Dir.</th><th>Route Plan</th><th>Appr. Order</th> </tr> <tr> <td>VST-3B-SFT</td><td>57.9</td><td>69.3</td><td>45.4</td><td>71.8</td><td>62.4</td><td>59.0</td><td>46.0</td><td>38.7</td><td>70.2</td> </tr> <tr> <td>VST-3B-RL</td><td>57.7</td><td>66.6</td><td>45.0</td><td>72.8</td><td>60.9</td><td>59.9</td><td>47.6</td><td>40.7</td><td>68.3</td> </tr> <tr> <td>VST-7B-SFT</td><td>60.6</td><td>72.0</td><td>44.4</td><td>74.3</td><td>68.3</td><td>59.7</td><td>55.8</td><td>44.9</td><td>65.2</td> </tr> <tr> <td>VST-7B-RL</td><td>61.2</td><td>71.6</td><td>43.8</td><td>75.5</td><td>69.2</td><td>60.0</td><td>55.6</td><td>44.3</td><td>69.2</td> </tr> </table> <h3>📈 SUN RGBD 3D Object Detection</h3> <table> <tr> <th>Methods</th><th>AP@15</th> </tr> <tr> <td>Seed1.5-VL</td><td>33.5</td> </tr> <tr> <td>Gemini-2.0-Pro</td><td>32.5</td> </tr> <tr> <td>Gemini Robotics-ER</td><td><b>48.3</b></td> </tr> <tr> <td>VST-3B-SFT</td><td>37.3</td> </tr> <tr> <td>VST-3B-RL</td><td>40.1</td> </tr> <tr> <td>VST-7B-SFT</td><td>41.6</td> </tr> <tr> <td>VST-7B-RL</td><td><b>44.2</b></td> </tr> </table> </details>

---

⚡ Getting Started

Training & Evaluation

SFT: Please follow assets/train.md to prepare env, data and train models.

RL: Please follow projects/spatial_rl/README.md to prepare env, data and train models using RL.

VLA: Please follow assets/vla.md to train vla models.

Evaluation: Please follow benchmark/README.md to evaluate models.

Cookbook

Cookbook	Description
scene understanding	Example for single image and multi-image inference
3d object detection	Example for 3D object detection

Using 🤗 Transformers to Chat

Install the inference dependency:

bashCopy
pip install transformers==4.57.0
pip install qwen-vl-utils

Then:

pythonCopy
import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

THINK_SYSTEM_PROMPT = "You are a helpful assistant. You should first think about the reasoning process in the mind and then provide the user with the answer. The reasoning process is enclosed within <think> </think> tags, i.e. <think> reasoning process here </think> answer here."
think_mesg = {
                "role": "system",
                "content": [{"type": "text", "text": THINK_SYSTEM_PROMPT}],
            }

enable_thinking=False

model_path="rayruiyang/VST-7B-RL"

# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

# default processer
processor = AutoProcessor.from_pretrained(model_path, min_pixels = 256*28*28, max_pixels=1280*28*28)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "http://images.cocodataset.org/train2017/000000075668.jpg",
            },
            {"type": "text", "text": "Consider the real-world 3D locations of the objects. Is the 'no motorcycle' sign directly above the red bus?"},
        ],
    }
]

if enable_thinking:
    messages.insert(0, think_mesg)


# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=1280)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text[0])

📊 Dataset Overview

🖼️ VST-Perception (VST-P)

• 4.1M samples across 19 tasks for supervised fine-tuning.
• Covers three primary vision scenarios: single-image, multi-image, and video.
• VLMs tuned on VST-P show strong improvements in spatial perception:
  • ~20% boost on CVBench-3D
  • ~5% increase on BLINK
  • ~16% gain on VSIBench

🧠 VST-Reasoning (VST-R)

• 135K samples, split into:
  • Reasoning steps (CoT): Teach models how to reason spatially.
  • Rule-checkable data: Used in online RL to further enhance reasoning skills.
• VLMs tuned on VST-R demonstrate:
  • 8.9% improvement on MMSI-Bench

---

There are 500K reproduced data points [rayruiyang/vst_500k] for academic purposes. You can download them:

shellCopy
python tools/download_hf_data.py --repo_id="rayruiyang/vst_500k" --local_dir $YOUR_LOCAL_PATH

[!NOTE]

• This data doesn't include the video files, please follow here to prepare video files.
• We use <|image_pad|> and <|video_pad|> as the image and video special token.

[Optional] You can parse the parquet data into a json file and raw images by:

bashCopy
python tools/parse_vst_500k.py --data_dir "$YOUR_LOCAL_PATH/vst_500k"

You will get the data:

textCopy
data/
├── images
├── vst_500k.json

📜 License

This project is licensed under the Apache License. See the LICENSE file for details.

The VST-3B model is fine-tuned from Qwen2.5VL-3B, its license is Qwen2.5VL-3B LICENSE.

Acknowledgement

Thanks for the projects: Qwen2.5VL, VeOmni, EasyR1, and VLMEvalKit.

If you find VST useful for your research or applications, please ⭐ star the repo or cite our work:

bibtexCopy
@article{vst,
  title={Visual Spatial Tuning},
  author={Rui Yang, Ziyu Zhu, Yanwei Li, Jingjia Huang, Shen Yan, Siyuan Zhou, Zhe Liu, Xiangtai Li, Shuangye Li, Wenqian Wang, Yi Lin, Hengshuang Zhao},
  journal={arXiv preprint arXiv:2511.05491},
  year={2025}
}