Open Qwen2VL

🚀 Open-Qwen2VL

Open-Qwen2VL is a multimodal model that accepts images and text as input and generates text as output. It effectively addresses the challenges in multimodal information processing and offers high - quality text output for image - text input scenarios.

🚀 Quick Start

Prerequisites

Ensure you have Python and pip installed on your system.

Installation

You can install Open-Qwen2VL using the following command:

pip install git+https://github.com/Victorwz/Open-Qwen2VL.git#subdirectory=prismatic-vlms

Inference

Here is an example of loading the model and performing inference:

import requests
import torch
from PIL import Image
from prismatic import load

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

# Load a pretrained VLM (either local path, or ID to auto-download from the HF Hub)
vlm = load("Open-Qwen2VL")
vlm.to(device, dtype=torch.bfloat16)

# Download an image and specify a prompt
image_url = "https://huggingface.co/adept/fuyu-8b/resolve/main/bus.png"
# image = Image.open(requests.get(image_url, stream=True).raw).convert("RGB")
image = [vlm.vision_backbone.image_transform(Image.open(requests.get(image_url, stream=True).raw).convert("RGB")).unsqueeze(0)]
user_prompt = "<image>\nDescribe the image."

# Generate!
generated_text = vlm.generate_batch(
    image,
    [user_prompt],
    do_sample=False,
    max_new_tokens=512,
    min_length=1,
)
print(generated_text[0])

The image caption results look like:

The image depicts a blue and orange bus parked on the side of a street. ...

✨ Features

• Multimodal Input: Accepts both images and text as input, enabling comprehensive multimodal information processing.
• Text Output: Generates high - quality text based on the input image and text.

📦 Model Information

📚 Documentation

This model is described in the paper Open-Qwen2VL: Compute-Efficient Pre-Training of Fully-Open Multimodal LLMs on Academic Resources. The code is available at https://github.com/Victorwz/Open-Qwen2VL.

Updates

• [4/1/2025] The codebase, model, data, and paper are released.

📄 License

The model is released under the cc license.

Acknowledgement

This work was partially supported by the BioPACIFIC Materials Innovation Platform of the National Science Foundation under Award No. DMR - 1933487

Citation

@article{Open-Qwen2VL,
    title={Open-Qwen2VL: Compute-Efficient Pre-Training of Fully-Open Multimodal LLMs on Academic Resources},
    author={Wang, Weizhi and Tian, Yu and Yang, Linjie and Wang, Heng and Yan, Xifeng},
    journal={arXiv preprint arXiv:2504.00595},
    year={2025}
  }