QWen-VL1.0

Qwen-VL

<br> <p align="center"> <img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/logo_vl.jpg" width="400"/> <p> <br> <p align="center"> Qwen-VL <a href="https://modelscope.cn/models/qwen/Qwen-VL/summary">🤖 <a> | <a href="https://huggingface.co/Qwen/Qwen-VL">🤗</a>&nbsp ｜ Qwen-VL-Chat <a href="https://modelscope.cn/models/qwen/Qwen-VL-Chat/summary">🤖 <a>| <a href="https://huggingface.co/Qwen/Qwen-VL-Chat">🤗</a>&nbsp ｜ Qwen-VL-Chat-Int4 <a href="https://huggingface.co/Qwen/Qwen-VL-Chat-Int4">🤗</a> <br> <a href="assets/wechat.png">WeChat</a>&nbsp&nbsp | &nbsp&nbsp<a href="https://discord.gg/z3GAxXZ9Ce">Discord</a>&nbsp&nbsp | &nbsp&nbsp<a href="https://modelscope.cn/studios/qwen/Qwen-VL-Chat-Demo/summary">Demo</a>&nbsp ｜ &nbsp<a href="https://arxiv.org/abs/2308.12966">Report</a> </p> <br>

Qwen-VL 是阿里云研发的大规模视觉语言模型（Large Vision Language Model, LVLM）。Qwen-VL 可以以图像、文本、检测框作为输入，并以文本和检测框作为输出。Qwen-VL 系列模型性能强大，具备多语言对话、多图交错对话等能力，并支持中文开放域定位和细粒度图像识别与理解。

Qwen-VL (Qwen Large Vision Language Model) is the visual multimodal version of the large model series, Qwen (abbr. Tongyi Qianwen), proposed by Alibaba Cloud. Qwen-VL accepts image, text, and bounding box as inputs, outputs text and bounding box. The features of Qwen-VL include:

目前，我们提供了Qwen-VL和Qwen-VL-Chat两个模型，分别为预训练模型和Chat模型。如果想了解更多关于模型的信息，请点击链接查看我们的技术备忘录。本仓库为Qwen-VL-Chat仓库。

We release Qwen-VL and Qwen-VL-Chat, which are pretrained model and Chat model respectively. For more details about Qwen-VL, please refer to our technical memo. This repo is the one for Qwen-VL. <br>

安装要求 (Requirements)

• python 3.8及以上版本
• pytorch 1.12及以上版本，推荐2.0及以上版本
• 建议使用CUDA 11.4及以上（GPU用户需考虑此选项）
• python 3.8 and above
• pytorch 1.12 and above, 2.0 and above are recommended
• CUDA 11.4 and above are recommended (this is for GPU users) <br>

快速开始 (Quickstart)

我们提供简单的示例来说明如何利用 🤗 Transformers 快速使用 Qwen-VL。

在开始前，请确保你已经配置好环境并安装好相关的代码包。最重要的是，确保你满足上述要求，然后安装相关的依赖库。

Below, we provide simple examples to show how to use Qwen-VL with 🤗 Transformers.

Before running the code, make sure you have setup the environment and installed the required packages. Make sure you meet the above requirements, and then install the dependent libraries.

pip install -r requirements.txt

接下来你可以开始使用Transformers来使用我们的模型。关于视觉模块的更多用法，请参考教程。

Now you can start with Transformers. More usage aboue vision encoder, please refer to tutorial.

🤗 Transformers

To use Qwen-VL for the inference, all you need to do is to input a few lines of codes as demonstrated below. However, please make sure that you are using the latest code.

from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation import GenerationConfig
import torch
torch.manual_seed(1234)

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-VL", trust_remote_code=True)

# use bf16
# model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL", device_map="auto", trust_remote_code=True, bf16=True).eval()
# use fp16
# model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL", device_map="auto", trust_remote_code=True, fp16=True).eval()
# use cpu only
# model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL", device_map="cpu", trust_remote_code=True).eval()
# use cuda device
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL", device_map="cuda", trust_remote_code=True).eval()

# Specify hyperparameters for generation (No need to do this if you are using transformers>=4.32.0)
# model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-VL", trust_remote_code=True)

query = tokenizer.from_list_format([
    {'image': 'https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg'},
    {'text': 'Generate the caption in English with grounding:'},
])
inputs = tokenizer(query, return_tensors='pt')
inputs = inputs.to(model.device)
pred = model.generate(**inputs)
response = tokenizer.decode(pred.cpu()[0], skip_special_tokens=False)
print(response)
# <img>https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg</img>Generate the caption in English with grounding:<ref> Woman</ref><box>(451,379),(731,806)</box> and<ref> her dog</ref><box>(219,424),(576,896)</box> playing on the beach<|endoftext|>
image = tokenizer.draw_bbox_on_latest_picture(response)
if image:
  image.save('2.jpg')
else:
  print("no box")

<p align="center"> <img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo_spotting_caption.jpg" width="500"/> <p> <br>

评测

我们从两个角度评测了两个模型的能力：

1. 在英文标准 Benchmark 上评测模型的基础任务能力。目前评测了四大类多模态任务：

   • Zero-shot Caption: 评测模型在未见过数据集上的零样本图片描述能力；
   • General VQA: 评测模型的通用问答能力，例如判断题、颜色、个数、类目等问答能力；
   • Text-based VQA：评测模型对于图片中文字相关的识别/问答能力，例如文档问答、图表问答、文字问答等；
   • Referring Expression Compression：评测模型给定物体描述画检测框的能力；

2. 试金石 (TouchStone)：为了评测模型整体的图文对话能力和人类对齐水平。我们为此构建了一个基于 GPT4 打分来评测 LVLM 模型的 Benchmark：TouchStone。在 TouchStone-v0.1 中：

   • 评测基准总计涵盖 300+张图片、800+道题目、27个类别。包括基础属性问答、人物地标问答、影视作品问答、视觉推理、反事实推理、诗歌创作、故事写作，商品比较、图片解题等尽可能广泛的类别。
   • 为了弥补目前 GPT4 无法直接读取图片的缺陷，我们给所有的带评测图片提供了人工标注的充分详细描述，并且将图片的详细描述、问题和模型的输出结果一起交给 GPT4 打分。
   • 评测同时包含英文版本和中文版本。

评测结果如下：

We evaluated the model's ability from two perspectives:

1. Standard Benchmarks: We evaluate the model's basic task capabilities on four major categories of multimodal tasks:

   • Zero-shot Caption: Evaluate model's zero-shot image captioning ability on unseen datasets;
   • General VQA: Evaluate the general question-answering ability of pictures, such as the judgment, color, number, category, etc;
   • Text-based VQA: Evaluate the model's ability to recognize text in pictures, such as document QA, chart QA, etc;
   • Referring Expression Comprehension: Evaluate the ability to localize a target object in an image described by a referring expression.

2. TouchStone: To evaluate the overall text-image dialogue capability and alignment level with humans, we have constructed a benchmark called TouchStone, which is based on scoring with GPT4 to evaluate the LVLM model.

   • The TouchStone benchmark covers a total of 300+ images, 800+ questions, and 27 categories. Such as attribute-based Q&A, celebrity recognition, writing poetry, summarizing multiple images, product comparison, math problem solving, etc;
   • In order to break the current limitation of GPT4 in terms of direct image input, TouchStone provides fine-grained image annotations by human labeling. These detailed annotations, along with the questions and the model's output, are then presented to GPT4 for scoring.
   • The benchmark includes both English and Chinese versions.

The results of the evaluation are as follows:

Qwen-VL outperforms current SOTA generalist models on multiple VL tasks and has a more comprehensive coverage in terms of capability range.

<p align="center"> <img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/radar.png" width="600"/> <p>

零样本图像描述 & 通用视觉问答 (Zero-shot Captioning & General VQA)

<table> <thead> <tr> <th rowspan="2">Model type</th> <th rowspan="2">Model</th> <th colspan="2">Zero-shot Captioning</th> <th colspan="5">General VQA</th> </tr> <tr> <th>NoCaps</th> <th>Flickr30K</th> <th>VQAv2^dev</th> <th>OK-VQA</th> <th>GQA</th> <th>SciQA-Img<br>(0-shot)</th> <th>VizWiz<br>(0-shot)</th> </tr> </thead> <tbody align="center"> <tr> <td rowspan="10">Generalist<br>Models</td> <td>Flamingo-9B</td> <td>-</td> <td>61.5</td> <td>51.8</td> <td>44.7</td> <td>-</td> <td>-</td> <td>28.8</td> </tr> <tr> <td>Flamingo-80B</td> <td>-</td> <td>67.2</td> <td>56.3</td> <td>50.6</td> <td>-</td> <td>-</td> <td>31.6</td> </tr> <tr> <td>Unified-IO-XL</td> <td>100.0</td> <td>-</td> <td>77.9</td> <td>54.0</td> <td>-</td> <td>-</td> <td>-</td> </tr> <tr> <td>Kosmos-1</td> <td>-</td> <td>67.1</td> <td>51.0</td> <td>-</td> <td>-</td> <td>-</td> <td>29.2</td> </tr> <tr> <td>Kosmos-2</td> <td>-</td> <td>66.7</td> <td>45.6</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> </tr> <tr> <td>BLIP-2 (Vicuna-13B)</td> <td>103.9</td> <td>71.6</td> <td>65.0</td> <td>45.9</td> <td>32.3</td> <td>61.0</td> <td>19.6</td> </tr> <tr> <td>InstructBLIP (Vicuna-13B)</td> <td><strong>121.9</strong></td> <td>82.8</td> <td>-</td> <td>-</td> <td>49.5</td> <td>63.1</td> <td>33.4</td> </tr> <tr> <td>Shikra (Vicuna-13B)</td> <td>-</td> <td>73.9</td> <td>77.36</td> <td>47.16</td> <td>-</td> <td>-</td> <td>-</td> </tr> <tr> <td><strong>Qwen-VL (Qwen-7B)</strong></td> <td>121.4</td> <td><b>85.8</b></td> <td><b>78.8</b></td> <td><b>58.6</b></td> <td><b>59.3</b></td> <td>67.1</td> <td>35.2</td> </tr> <!-- <tr> <td>Qwen-VL (4-shot)</td> <td>-</td> <td>-</td> <td>-</td> <td>63.6</td> <td>-</td> <td>-</td> <td>39.1</td> </tr> --> <tr> <td>Qwen-VL-Chat</td> <td>120.2</td> <td>81.0</td> <td>78.2</td> <td>56.6</td> <td>57.5</td> <td><b>68.2</b></td> <td><b>38.9</b></td> </tr> <!-- <tr> <td>Qwen-VL-Chat (4-shot)</td> <td>-</td> <td>-</td> <td>-</td> <td>60.6</td> <td>-</td> <td>-</td> <td>44.45</td> </tr> --> <tr> <td>Previous SOTA<br>(Per Task Fine-tuning)</td> <td>-</td> <td>127.0<br>(PALI-17B)</td> <td>84.5<br>(InstructBLIP<br>-FlanT5-XL)</td> <td>86.1<br>(PALI-X<br>-55B)</td> <td>66.1<br>(PALI-X<br>-55B)</td> <td>72.1<br>(CFR)</td> <td>92.53<br>(LLaVa+<br>GPT-4)</td> <td>70.9<br>(PALI-X<br>-55B)</td> </tr> </tbody> </table>

• 在 Zero-shot Caption 中，Qwen-VL 在 Flickr30K 数据集上取得了 SOTA 的结果，并在 Nocaps 数据集上取得了和 InstructBlip 可竞争的结果。
• 在 General VQA 中，Qwen-VL 取得了 LVLM 模型同等量级和设定下 SOTA 的结果。
• For zero-shot image captioning, Qwen-VL achieves the SOTA on Flickr30K and competitive results on Nocaps with InstructBlip.
• For general VQA, Qwen-VL achieves the SOTA under the same generalist LVLM scale settings.

文本导向的视觉问答 (Text-oriented VQA)

<table> <thead> <tr> <th>Model type</th> <th>Model</th> <th>TextVQA</th> <th>DocVQA</th> <th>ChartQA</th> <th>AI2D</th> <th>OCR-VQA</th> </tr> </thead> <tbody align="center"> <tr> <td rowspan="5">Generalist Models</td> <td>BLIP-2 (Vicuna-13B)</td> <td>42.4</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> </tr> <tr> <td>InstructBLIP (Vicuna-13B)</td> <td>50.7</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> </tr> <tr> <td>mPLUG-DocOwl (LLaMA-7B)</td> <td>52.6</td> <td>62.2</td> <td>57.4</td> <td>-</td> <td>-</td> </tr> <tr> <td>Pic2Struct-Large (1.3B)</td> <td>-</td> <td><b>76.6</b></td> <td>58.6</td> <td>42.1</td> <td>71.3</td> </tr> <tr> <td>Qwen-VL (Qwen-7B)</td> <td><b>63.8</b></td> <td>65.1</td> <td><b>65.7</b></td> <td><b>62.3</b></td> <td><b>75.7</b></td> </tr> <tr> <td>Specialist SOTAs<br>(Specialist/Finetuned)</td> <td>PALI-X-55B (Single-task FT)<br>(Without OCR Pipeline)</td> <td>71.44</td> <td>80.0</td> <td>70.0</td> <td>81.2</td> <td>75.0</td> </tr> </tbody> </table>

• 在文字相关的识别/问答评测上，取得了当前规模下通用 LVLM 达到的最好结果。
• 分辨率对上述某几个评测非常重要，大部分 224 分辨率的开源 LVLM 模型无法完成以上评测，或只能通过切图的方式解决。Qwen-VL 将分辨率提升到 448，可以直接以端到端的方式进行以上评测。Qwen-VL 在很多任务上甚至超过了 1024 分辨率的 Pic2Struct-Large 模型。
• In text-related recognition/QA evaluation, Qwen-VL achieves the SOTA under the generalist LVLM scale settings.
• Resolution is important for several above evaluations. While most open-source LVLM models with 224 resolution are incapable of these evaluations or can only solve these by cutting images, Qwen-VL scales the resolution to 448 so that it can be evaluated end-to-end. Qwen-VL even outperforms Pic2Struct-Large models of 1024 resolution on some tasks.

细粒度视觉定位 (Referring Expression Comprehension)

<table> <thead> <tr> <th rowspan="2">Model type</th> <th rowspan="2">Model</th> <th colspan="3">RefCOCO</th> <th colspan="3">RefCOCO+</th> <th colspan="2">RefCOCOg</th> <th>GRIT</th> </tr> <tr> <th>val</th> <th>test-A</th> <th>test-B</th> <th>val</th> <th>test-A</th> <th>test-B</th> <th>val-u</th> <th>test-u</th> <th>refexp</th> </tr> </thead> <tbody align="center"> <tr> <td rowspan="8">Generalist Models</td> <td>GPV-2</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>51.50</td> </tr> <tr> <td>OFA-L*</td> <td>79.96</td> <td>83.67</td> <td>76.39</td> <td>68.29</td> <td>76.00</td> <td>61.75</td> <td>67.57</td> <td>67.58</td> <td>61.70</td> </tr> <tr> <td>Unified-IO</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td><b>78.61</b></td> </tr> <tr> <td>VisionLLM-H</td> <td></td> <td>86.70</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> </tr> <tr> <td>Shikra-7B</td> <td>87.01</td> <td>90.61</td> <td>80.24 </td> <td>81.60</td> <td>87.36</td> <td>72.12</td> <td>82.27</td> <td>82.19</td> <td>69.34</td> </tr> <tr> <td>Shikra-13B</td> <td>87.83 </td> <td>91.11</td> <td>81.81</td> <td>82.89</td> <td>87.79</td> <td>74.41</td> <td>82.64</td> <td>83.16</td> <td>69.03</td> </tr> <tr> <td>Qwen-VL-7B</td> <td><b>89.36</b></td> <td>92.26</td> <td><b>85.34</b></td> <td><b>83.12</b></td> <td>88.25</td> <td><b>77.21</b></td> <td>85.58</td> <td>85.48</td> <td>78.22</td> </tr> <tr> <td>Qwen-VL-7B-Chat</td> <td>88.55</td> <td><b>92.27</b></td> <td>84.51</td> <td>82.82</td> <td><b>88.59</b></td> <td>76.79</td> <td><b>85.96</b></td> <td><b>86.32</b></td> <td>-</td> <tr> <td rowspan="3">Specialist SOTAs<br>(Specialist/Finetuned)</td> <td>G-DINO-L</td> <td>90.56&nbsp;&nbsp;</td> <td>93.19</td> <td>88.24</td> <td>82.75</td> <td>88.95</td> <td>75.92</td> <td>86.13</td> <td>87.02</td> <td>-</td> </tr> <tr> <td>UNINEXT-H</td> <td>92.64 </td> <td>94.33</td> <td>91.46</td> <td>85.24</td> <td>89.63</td> <td>79.79</td> <td>88.73</td> <td>89.37</td> <td>-</td> </tr> <tr> <td>ONE-PEACE</td> <td>92.58 </td> <td>94.18</td> <td>89.26</td> <td>88.77</td> <td>92.21</td> <td>83.23</td> <td>89.22</td> <td>89.27</td> <td>-</td> </tr> </tbody> </table>

• 在定位任务上，Qwen-VL 全面超过 Shikra-13B，取得了目前 Generalist LVLM 模型上在 Refcoco 上的 SOTA。
• Qwen-VL 并没有在任何中文定位数据上训练过，但通过中文 Caption 数据和 英文 Grounding 数据的训练，可以 Zero-shot 泛化出中文 Grounding 能力。

我们提供了以上所有评测脚本以供复现我们的实验结果。请阅读 eval/EVALUATION.md 了解更多信息。

• Qwen-VL achieves the SOTA in all above referring expression comprehension benchmarks.
• Qwen-VL has not been trained on any Chinese grounding data, but it can still generalize to the Chinese Grounding tasks in a zero-shot way by training Chinese Caption data and English Grounding data.

We provide all of the above evaluation scripts for reproducing our experimental results. Please read eval/EVALUATION.md for more information.

闲聊能力测评 (Chat Evaluation)

TouchStone 是一个基于 GPT4 打分来评测 LVLM 模型的图文对话能力和人类对齐水平的基准。它涵盖了 300+张图片、800+道题目、27个类别，包括基础属性、人物地标、视觉推理、诗歌创作、故事写作、商品比较、图片解题等尽可能广泛的类别。关于 TouchStone 的详细介绍，请参考touchstone/README_CN.md了解更多信息。

TouchStone is a benchmark based on scoring with GPT4 to evaluate the abilities of the LVLM model on text-image dialogue and alignment levels with humans. It covers a total of 300+ images, 800+ questions, and 27 categories, such as attribute-based Q&A, celebrity recognition, writing poetry, summarizing multiple images, product comparison, math problem solving, etc. Please read touchstone/README_CN.md for more information.

英语 (English)

中文 (Chinese)

Qwen-VL-Chat 模型在中英文的对齐评测中均取得当前 LVLM 模型下的最好结果。

Qwen-VL-Chat has achieved the best results in both Chinese and English alignment evaluation. <br>

常见问题 (FAQ)

如遇到问题，敬请查阅 FAQ以及issue区，如仍无法解决再提交issue。

If you meet problems, please refer to FAQ and the issues first to search a solution before you launch a new issue. <br>

使用协议 (License Agreement)

研究人员与开发者可使用Qwen-VL和Qwen-VL-Chat或进行二次开发。我们同样允许商业使用，具体细节请查看LICENSE。如需商用，请填写问卷申请。

Researchers and developers are free to use the codes and model weights of both Qwen-VL and Qwen-VL-Chat. We also allow their commercial use. Check our license at LICENSE for more details. <br>

引用 (Citation)

如果你觉得我们的论文和代码对你的研究有帮助，请考虑:star: 和引用 :pencil: :)

If you find our paper and code useful in your research, please consider giving a star :star: and citation :pencil: :)

@article{Qwen-VL,
  title={Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilities},
  author={Bai, Jinze and Bai, Shuai and Yang, Shusheng and Wang, Shijie and Tan, Sinan and Wang, Peng and Lin, Junyang and Zhou, Chang and Zhou, Jingren},
  journal={arXiv preprint arXiv:2308.12966},
  year={2023}
}

<br>

联系我们 (Contact Us)

如果你想给我们的研发团队和产品团队留言，请通过邮件（[email protected]）联系我们。

If you are interested to leave a message to either our research team or product team, feel free to send an email to [email protected].