VLM-FO1

VLM-FO1: Bridging the Gap Between High-Level Reasoning and Fine-Grained Perception

VLM-FO1 endows pre-trained VLMs with superior fine-grained perception without compromising their inherent high-level reasoning and general understanding capabilities. It operates as a plug-and-play module that can be integrated with any existing VLM, establishing an effective and flexible paradigm for building the next generation of perception-aware models.

VLM-FO1 excels at a wide range of fine-grained perception tasks, including Object Grounding, Region Generative Understanding, Visual Region Reasoning, and more.

For more details, check out our paper

📰 Update News

• 2026-03-12 — We are excited to announce that our models are now available on ModelScope!
• 2025-11-28 — We released a new video tracking demo that combines SAM3 video tracking with VLM-FO1, enabling interactive object detection and tracking across video frames with complex natural language prompts.
• 2025-11-21 — We released a new demo that integrates SAM3 with VLM-FO1, enabling stronger segmentation fidelity and more reliable detection under complex, compositional label prompts. Check this section Inference with SAM3 + VLM-FO1 Gradio Demo about how to run it. Or try the huggingface demo

✨ Key Features

🧩 Plug-and-Play Modularity: Our framework is designed as a set of enhancement modules that can be seamlessly integrated with any pre-trained VLM, preserving its original weights and capabilities.

🧠 Hybrid Region Encoder (HFRE): We introduce a novel Dual-Vision Encoder architecture that fuses semantic-rich features with perception-enhanced features, creating powerful region tokens that capture both high-level meaning and fine-grained spatial detail.

🎯 State-of-the-Art Performance: VLM-FO1 achieves SOTA results across a diverse suite of benchmarks.

✅ Preserves General Abilities: Our two-stage training strategy ensures that fine-grained perception is gained without causing catastrophic forgetting of the base model’s powerful general visual understanding abilities.

📝 ToDo

• Release Gradio demo

🔗 Related works

If you are interested in our research, we welcome you to explore our other wonderful projects.

• 🔆 Om AI Lab Blogs
• 🔆 VLM-R1: A stable and generalizable R1-style Large Vision-Language Model
• 🔆 OmAgent: Build Multimodal Language Agents with Ease
• 🔆 OmDet-Turbo: Real-time and accurate open-vocabulary end-to-end object detection
• 🔆 ZoomEye: Enhancing Multimodal LLMs with Human-Like Zooming Capabilities through Tree-Based Image Exploration

📋 Table of Contents

• Installation
• Quick Start
  • Inference with Provided Bounding Boxes
  • Inference with Object Detector
  • Inference with Gradio Demo
  • Inference with SAM3 + VLM-FO1 Gradio Demo
  • Inference with SAM3 + VLM-FO1 Video Tracking Gradio Demo
• Task Templates
• Evaluation
• Citation
• Acknowledgements

🛠️ Installation

Requirements

• Python 3.10+
• PyTorch (GPU recommended). Please install the CUDA-enabled build that matches your system.
• Linux is the primary tested platform.

Setup

git clone https://github.com/om-ai-lab/VLM-FO1.git
cd VLM-FO1
pip install -r requirements.txt

If you use conda:

conda create -n vlm-fo1 python=3.10 -y
conda activate vlm-fo1
pip install -r requirements.txt

🚀 Quick Start

1) Download Model Checkpoints

Download the pre-trained VLM-FO1 checkpoints from Hugging Face and place them under resources/:

• Pre-trained model: Hugging Face link

2) Inference with Provided Bounding Boxes

Use inference.py to run VLM-FO1 with provided bounding boxes.

Run:

python inference.py

The visualization with predicted boxes will be saved to demo/vlm_fo1_result.jpg.

3) Inference with Object Detector

Option A: Inference with UPN

Note: Due to company policy, we are unable to release the object detector OPN referenced in our paper. Instead, we provide integration with UPN (from ChatRex), which offers similar functionality. You are also free to use any object detector of your choice by preparing bounding box proposals compatible with our pipeline.

Before running, you need to install UPN’s requirements and build the ops extension:

cd detect_tools/upn
pip install -r requirements.txt
cd ops
pip install -v -e .

Then, download the UPN checkpoint from this link and place it in your resources/ folder:

wget https://github.com/IDEA-Research/ChatRex/releases/download/upn-large/upn_large.pth -P resources/

Use scripts/inference_with_upn.py to run FO1 with UPN-generated proposals.

Run:

python scripts/inference_with_upn.py

The script will:

• Load UPN and produce fine-grained object proposals
• Filter proposals by score
• Pass top proposals to FO1 for reasoning
• Save the visualization to demo/vlm_fo1_result.jpg

Option B: Inference with SAM3

We also provide integration with SAM3, which offers high-quality open-vocabulary segmentation. This integration leverages SAM3 to generate object proposals based on text prompts, which are then refined and selected by VLM-FO1.

1. Follow the official SAM3 setup guide to install sam3 and request access to the checkpoints.
2. Download the SAM3 checkpoint (e.g., sam3.pt) and place it in resources/sam3/.

Use scripts/inference_with_sam3.py to run FO1 with SAM3-generated proposals.

Run:

python scripts/inference_with_sam3.py

The script will:

• Load SAM3 and generate object proposals based on a text prompt (e.g., “the ball nearest to the bear”)
• Pass top proposals (bounding boxes) to FO1 for reasoning and filtering
• Save the visualization to demo/vlm_fo1_result_with_sam3.jpg

4) Inference with Gradio Demo

Run:

python demo/gradio_demo.py

5) Inference with SAM3 + VLM-FO1 Gradio Demo

1. Follow the official SAM3 setup guide to install sam3 and request access to the checkpoints from Hugging Face.
2. Launch the joint demo:

python demo/gradio_demo_with_sam3.py

This script pipes SAM3’s text-conditioned segmentation into VLM-FO1’s complex label reasoning for more reliable detection and segmentation on compositional prompts.

6) Inference with SAM3 + VLM-FO1 Video Tracking Gradio Demo

Launch the video tracking demo:

python demo/gradio_demo_with_sam3_video_tracking.py

This demo extends the capabilities to video:

1. Frame Selection: Pick a keyframe from your video.
2. Detection: Use VLM-FO1 + SAM3 to identify objects with complex text prompts on the selected frame.
3. Tracking: Propagate the segmentation masks forward through the video using SAM3’s tracker.

📝 Task Templates

We provide task templates to facilitate tasks beyond detection (e.g., counting, grounding, and other vision-language reasoning tasks). See vlm_fo1/task_templates for examples. You can adapt the prompts or message construction to your specific application.

Evaluation

Model performance when integrated with UPN is as follows:

Model	COCO mAP	CountBench Accuracy	Pixmo-Count Accuracy	HumanRef DF1/P/R	LVIS SS/S-IoU	PACO SS/S-IoU	COCOText
VLM-FO1-3B-v01 (FO1 + UPN)	44.4	89.73%	85.07%	82.6/86.8/83.5	92.5/87	88.1/77.9	57.1

For benchmarking FO1 with UPN on your datasets, you can adapt the inference pipeline to export predictions and compare against ground-truth. We also include evaluation utilities under evaluation/ for specific tasks.

• CountBench/Pixmo-Count evaluation: evaluation/eval_countbench.py
• COCO evaluation: evaluation/eval_coco.py (this script will output a prediction json file which you can use with the standard COCO evaluation tool to obtain scores)

📄 Citation

If you find VLM-FO1 useful in your research or applications, please cite our work:

@article{liu2025vlm,
  title={VLM-FO1: Bridging the Gap Between High-Level Reasoning and Fine-Grained Perception in VLMs},
  author={Liu, Peng and Shen, Haozhan and Fang, Chunxin and Sun, Zhicheng and Liao, Jiajia and Zhao, Tiancheng},
  journal={arXiv preprint arXiv:2509.25916},
  year={2025}
}

🙏 Acknowledgements

• Dataset images used above are from public benchmarks (Pixmo-Count, CountBenchQA), credited to their respective authors.
• Thanks to the numerous researchers and developers who have contributed to the datasets used in our training.