Visual Spatial Tuning: VST-3B-SFT

This model is described in the paper Visual Spatial Tuning.

TL;DR: VST is a comprehensive framework designed to cultivate Vision-Language Models (VLMs) with human-like visuospatial abilities—from spatial perception to advanced reasoning.

💡 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 knowledge, 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 spatial tuning, paving the way for more physically grounded AI.

📈 Spatial & General Benchmarks

Models	CV	3DSR	MMSI	BLINK	VSI	MMStar	MMB	RealworldQA	MMMU	OCRB	AI2D
VST-3B-SFT	84.4	54.1	30.2	59.1	57.9	58.0	80.9	68.4	45.2	83.7	82.5
VST-3B-RL	84.2	56.5	31.3	57.2	57.7	58.9	80.5	68.5	49.8	80.9	82.4
VST-7B-SFT	85.5	54.6	32.0	62.1	60.6	63.1	83.3	72.2	50.6	85.5	84.9
VST-7B-RL	86.5	60.1	34.8	62.6	61.2	63.5	83.0	68.5	49.4	86.1	83.5

📈 VSIBench

Methods	Avg.	Obj. Count	Abs. Dist.	Obj. Size	Room Size	Rel. Dist	Rel. Dir.	Route Plan	Appr. Order
VST-3B-SFT	57.9	69.3	45.4	71.8	62.4	59.0	46.0	38.7	70.2
VST-3B-RL	57.7	66.6	45.0	72.8	60.9	59.9	47.6	40.7	68.3
VST-7B-SFT	60.6	72.0	44.4	74.3	68.3	59.7	55.8	44.9	65.2
VST-7B-RL	61.2	71.6	43.8	75.5	69.2	60.0	55.6	44.3	69.2

Quickstart

pip install transformers
# It's highly recommanded to use `[decord]` feature for faster video loading.
pip install qwen-vl-utils

Using 🤗 Transformers to Chat

Here we show a code snippet to show you how to use the chat model with transformers and qwen_vl_utils:

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

model_path="rayruiyang/VST-3B-SFT"

# 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/train2017/000000039685.jpg",
            },
            {"type": "text", "text": "Consider the real-world 3D locations of the objects. Is the flag directly underneath the airplane?"},
        ],
    }
]

# 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=128)
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])

Citation

If you find our work helpful, feel free to give us a cite.

@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}
}