cydxg commited on
Commit
c814b23
·
verified ·
1 Parent(s): 25e500b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +525 -3
README.md CHANGED
@@ -1,3 +1,525 @@
1
- ---
2
- license: mit
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ pipeline_tag: image-text-to-text
6
+ tags:
7
+ - multimodal
8
+ library_name: transformers
9
+ ---
10
+ This is the OpenVINO accelerate version for multimodal Qwen2-VL-2B-Instruct.
11
+ To use this model, download all files and place the file as follow:
12
+ ```
13
+ .
14
+ │ gradio_helper.py
15
+ │ ov_qwen2_vl.py
16
+ │ qwen2-build.py
17
+ │ qwen2vl.ipynb
18
+ ├─Qwen2-VL-2B-Instruct
19
+ │ added_tokens.json
20
+ │ chat_template.json
21
+ │ config.json
22
+ │ merges.txt
23
+ │ openvino_language_model.bin
24
+ │ openvino_language_model.xml
25
+ │ openvino_text_embeddings_model.bin
26
+ │ openvino_text_embeddings_model.xml
27
+ │ openvino_vision_embeddings_merger_model.bin
28
+ │ openvino_vision_embeddings_merger_model.xml
29
+ │ openvino_vision_embeddings_model.bin
30
+ │ openvino_vision_embeddings_model.xml
31
+ │ preprocessor_config.json
32
+ │ special_tokens_map.json
33
+ │ tokenizer.json
34
+ │ tokenizer_config.json
35
+ │ vocab.json
36
+ ```
37
+ then run the qwen2vl.ipynb on your Jupyter Notebook.
38
+
39
+ Below is the original Model Card:
40
+
41
+ # Qwen2-VL-2B-Instruct
42
+
43
+ ## Introduction
44
+
45
+ We're excited to unveil **Qwen2-VL**, the latest iteration of our Qwen-VL model, representing nearly a year of innovation.
46
+
47
+ ### What’s New in Qwen2-VL?
48
+
49
+ #### Key Enhancements:
50
+
51
+
52
+ * **SoTA understanding of images of various resolution & ratio**: Qwen2-VL achieves state-of-the-art performance on visual understanding benchmarks, including MathVista, DocVQA, RealWorldQA, MTVQA, etc.
53
+
54
+ * **Understanding videos of 20min+**: Qwen2-VL can understand videos over 20 minutes for high-quality video-based question answering, dialog, content creation, etc.
55
+
56
+ * **Agent that can operate your mobiles, robots, etc.**: with the abilities of complex reasoning and decision making, Qwen2-VL can be integrated with devices like mobile phones, robots, etc., for automatic operation based on visual environment and text instructions.
57
+
58
+ * **Multilingual Support**: to serve global users, besides English and Chinese, Qwen2-VL now supports the understanding of texts in different languages inside images, including most European languages, Japanese, Korean, Arabic, Vietnamese, etc.
59
+
60
+
61
+ #### Model Architecture Updates:
62
+
63
+ * **Naive Dynamic Resolution**: Unlike before, Qwen2-VL can handle arbitrary image resolutions, mapping them into a dynamic number of visual tokens, offering a more human-like visual processing experience.
64
+
65
+ <p align="center">
66
+ <img src="https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen2-VL/qwen2_vl.jpg" width="80%"/>
67
+ <p>
68
+
69
+ * **Multimodal Rotary Position Embedding (M-ROPE)**: Decomposes positional embedding into parts to capture 1D textual, 2D visual, and 3D video positional information, enhancing its multimodal processing capabilities.
70
+
71
+ <p align="center">
72
+ <img src="http://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen2-VL/mrope.png" width="80%"/>
73
+ <p>
74
+
75
+ We have three models with 2, 7 and 72 billion parameters. This repo contains the instruction-tuned 2B Qwen2-VL model. For more information, visit our [Blog](https://qwenlm.github.io/blog/qwen2-vl/) and [GitHub](https://github.com/QwenLM/Qwen2-VL).
76
+
77
+
78
+
79
+ ## Evaluation
80
+
81
+ ### Image Benchmarks
82
+
83
+ | Benchmark | InternVL2-2B | MiniCPM-V 2.0 | **Qwen2-VL-2B** |
84
+ | :--- | :---: | :---: | :---: |
85
+ | MMMU<sub>val</sub> | 36.3 | 38.2 | **41.1** |
86
+ | DocVQA<sub>test</sub> | 86.9 | - | **90.1** |
87
+ | InfoVQA<sub>test</sub> | 58.9 | - | **65.5** |
88
+ | ChartQA<sub>test</sub> | **76.2** | - | 73.5 |
89
+ | TextVQA<sub>val</sub> | 73.4 | - | **79.7** |
90
+ | OCRBench | 781 | 605 | **794** |
91
+ | MTVQA | - | - | **20.0** |
92
+ | VCR<sub>en easy</sub> | - | - | **81.45**
93
+ | VCR<sub>zh easy</sub> | - | - | **46.16**
94
+ | RealWorldQA | 57.3 | 55.8 | **62.9** |
95
+ | MME<sub>sum</sub> | **1876.8** | 1808.6 | 1872.0 |
96
+ | MMBench-EN<sub>test</sub> | 73.2 | 69.1 | **74.9** |
97
+ | MMBench-CN<sub>test</sub> | 70.9 | 66.5 | **73.5** |
98
+ | MMBench-V1.1<sub>test</sub> | 69.6 | 65.8 | **72.2** |
99
+ | MMT-Bench<sub>test</sub> | - | - | **54.5** |
100
+ | MMStar | **49.8** | 39.1 | 48.0 |
101
+ | MMVet<sub>GPT-4-Turbo</sub> | 39.7 | 41.0 | **49.5** |
102
+ | HallBench<sub>avg</sub> | 38.0 | 36.1 | **41.7** |
103
+ | MathVista<sub>testmini</sub> | **46.0** | 39.8 | 43.0 |
104
+ | MathVision | - | - | **12.4** |
105
+
106
+ ### Video Benchmarks
107
+
108
+ | Benchmark | **Qwen2-VL-2B** |
109
+ | :--- | :---: |
110
+ | MVBench | **63.2** |
111
+ | PerceptionTest<sub>test</sub> | **53.9** |
112
+ | EgoSchema<sub>test</sub> | **54.9** |
113
+ | Video-MME<sub>wo/w subs</sub> | **55.6**/**60.4** |
114
+
115
+
116
+ ## Requirements
117
+ The code of Qwen2-VL has been in the latest Hugging face transformers and we advise you to build from source with command `pip install git+https://github.com/huggingface/transformers`, or you might encounter the following error:
118
+ ```
119
+ KeyError: 'qwen2_vl'
120
+ ```
121
+
122
+ ## Quickstart
123
+ We offer a toolkit to help you handle various types of visual input more conveniently. This includes base64, URLs, and interleaved images and videos. You can install it using the following command:
124
+
125
+ ```bash
126
+ pip install qwen-vl-utils
127
+ ```
128
+
129
+ Here we show a code snippet to show you how to use the chat model with `transformers` and `qwen_vl_utils`:
130
+
131
+ ```python
132
+ from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
133
+ from qwen_vl_utils import process_vision_info
134
+
135
+ # default: Load the model on the available device(s)
136
+ model = Qwen2VLForConditionalGeneration.from_pretrained(
137
+ "Qwen/Qwen2-VL-2B-Instruct", torch_dtype="auto", device_map="auto"
138
+ )
139
+
140
+ # We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
141
+ # model = Qwen2VLForConditionalGeneration.from_pretrained(
142
+ # "Qwen/Qwen2-VL-2B-Instruct",
143
+ # torch_dtype=torch.bfloat16,
144
+ # attn_implementation="flash_attention_2",
145
+ # device_map="auto",
146
+ # )
147
+
148
+ # default processer
149
+ processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
150
+
151
+ # The default range for the number of visual tokens per image in the model is 4-16384. You can set min_pixels and max_pixels according to your needs, such as a token count range of 256-1280, to balance speed and memory usage.
152
+ # min_pixels = 256*28*28
153
+ # max_pixels = 1280*28*28
154
+ # processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels)
155
+
156
+ messages = [
157
+ {
158
+ "role": "user",
159
+ "content": [
160
+ {
161
+ "type": "image",
162
+ "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
163
+ },
164
+ {"type": "text", "text": "Describe this image."},
165
+ ],
166
+ }
167
+ ]
168
+
169
+ # Preparation for inference
170
+ text = processor.apply_chat_template(
171
+ messages, tokenize=False, add_generation_prompt=True
172
+ )
173
+ image_inputs, video_inputs = process_vision_info(messages)
174
+ inputs = processor(
175
+ text=[text],
176
+ images=image_inputs,
177
+ videos=video_inputs,
178
+ padding=True,
179
+ return_tensors="pt",
180
+ )
181
+ inputs = inputs.to("cuda")
182
+
183
+ # Inference: Generation of the output
184
+ generated_ids = model.generate(**inputs, max_new_tokens=128)
185
+ generated_ids_trimmed = [
186
+ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
187
+ ]
188
+ output_text = processor.batch_decode(
189
+ generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
190
+ )
191
+ print(output_text)
192
+ ```
193
+ <details>
194
+ <summary>Without qwen_vl_utils</summary>
195
+
196
+ ```python
197
+ from PIL import Image
198
+ import requests
199
+ import torch
200
+ from torchvision import io
201
+ from typing import Dict
202
+ from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
203
+
204
+ # Load the model in half-precision on the available device(s)
205
+ model = Qwen2VLForConditionalGeneration.from_pretrained(
206
+ "Qwen/Qwen2-VL-2B-Instruct", torch_dtype="auto", device_map="auto"
207
+ )
208
+ processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
209
+
210
+ # Image
211
+ url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
212
+ image = Image.open(requests.get(url, stream=True).raw)
213
+
214
+ conversation = [
215
+ {
216
+ "role": "user",
217
+ "content": [
218
+ {
219
+ "type": "image",
220
+ },
221
+ {"type": "text", "text": "Describe this image."},
222
+ ],
223
+ }
224
+ ]
225
+
226
+
227
+ # Preprocess the inputs
228
+ text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
229
+ # Excepted output: '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe this image.<|im_end|>\n<|im_start|>assistant\n'
230
+
231
+ inputs = processor(
232
+ text=[text_prompt], images=[image], padding=True, return_tensors="pt"
233
+ )
234
+ inputs = inputs.to("cuda")
235
+
236
+ # Inference: Generation of the output
237
+ output_ids = model.generate(**inputs, max_new_tokens=128)
238
+ generated_ids = [
239
+ output_ids[len(input_ids) :]
240
+ for input_ids, output_ids in zip(inputs.input_ids, output_ids)
241
+ ]
242
+ output_text = processor.batch_decode(
243
+ generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
244
+ )
245
+ print(output_text)
246
+ ```
247
+ </details>
248
+
249
+ <details>
250
+ <summary>Multi image inference</summary>
251
+
252
+ ```python
253
+ # Messages containing multiple images and a text query
254
+ messages = [
255
+ {
256
+ "role": "user",
257
+ "content": [
258
+ {"type": "image", "image": "file:///path/to/image1.jpg"},
259
+ {"type": "image", "image": "file:///path/to/image2.jpg"},
260
+ {"type": "text", "text": "Identify the similarities between these images."},
261
+ ],
262
+ }
263
+ ]
264
+
265
+ # Preparation for inference
266
+ text = processor.apply_chat_template(
267
+ messages, tokenize=False, add_generation_prompt=True
268
+ )
269
+ image_inputs, video_inputs = process_vision_info(messages)
270
+ inputs = processor(
271
+ text=[text],
272
+ images=image_inputs,
273
+ videos=video_inputs,
274
+ padding=True,
275
+ return_tensors="pt",
276
+ )
277
+ inputs = inputs.to("cuda")
278
+
279
+ # Inference
280
+ generated_ids = model.generate(**inputs, max_new_tokens=128)
281
+ generated_ids_trimmed = [
282
+ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
283
+ ]
284
+ output_text = processor.batch_decode(
285
+ generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
286
+ )
287
+ print(output_text)
288
+ ```
289
+ </details>
290
+
291
+ <details>
292
+ <summary>Video inference</summary>
293
+
294
+ ```python
295
+ # Messages containing a images list as a video and a text query
296
+ messages = [
297
+ {
298
+ "role": "user",
299
+ "content": [
300
+ {
301
+ "type": "video",
302
+ "video": [
303
+ "file:///path/to/frame1.jpg",
304
+ "file:///path/to/frame2.jpg",
305
+ "file:///path/to/frame3.jpg",
306
+ "file:///path/to/frame4.jpg",
307
+ ],
308
+ "fps": 1.0,
309
+ },
310
+ {"type": "text", "text": "Describe this video."},
311
+ ],
312
+ }
313
+ ]
314
+ # Messages containing a video and a text query
315
+ messages = [
316
+ {
317
+ "role": "user",
318
+ "content": [
319
+ {
320
+ "type": "video",
321
+ "video": "file:///path/to/video1.mp4",
322
+ "max_pixels": 360 * 420,
323
+ "fps": 1.0,
324
+ },
325
+ {"type": "text", "text": "Describe this video."},
326
+ ],
327
+ }
328
+ ]
329
+
330
+ # Preparation for inference
331
+ text = processor.apply_chat_template(
332
+ messages, tokenize=False, add_generation_prompt=True
333
+ )
334
+ image_inputs, video_inputs = process_vision_info(messages)
335
+ inputs = processor(
336
+ text=[text],
337
+ images=image_inputs,
338
+ videos=video_inputs,
339
+ padding=True,
340
+ return_tensors="pt",
341
+ )
342
+ inputs = inputs.to("cuda")
343
+
344
+ # Inference
345
+ generated_ids = model.generate(**inputs, max_new_tokens=128)
346
+ generated_ids_trimmed = [
347
+ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
348
+ ]
349
+ output_text = processor.batch_decode(
350
+ generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
351
+ )
352
+ print(output_text)
353
+ ```
354
+ </details>
355
+
356
+ <details>
357
+ <summary>Batch inference</summary>
358
+
359
+ ```python
360
+ # Sample messages for batch inference
361
+ messages1 = [
362
+ {
363
+ "role": "user",
364
+ "content": [
365
+ {"type": "image", "image": "file:///path/to/image1.jpg"},
366
+ {"type": "image", "image": "file:///path/to/image2.jpg"},
367
+ {"type": "text", "text": "What are the common elements in these pictures?"},
368
+ ],
369
+ }
370
+ ]
371
+ messages2 = [
372
+ {"role": "system", "content": "You are a helpful assistant."},
373
+ {"role": "user", "content": "Who are you?"},
374
+ ]
375
+ # Combine messages for batch processing
376
+ messages = [messages1, messages1]
377
+
378
+ # Preparation for batch inference
379
+ texts = [
380
+ processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True)
381
+ for msg in messages
382
+ ]
383
+ image_inputs, video_inputs = process_vision_info(messages)
384
+ inputs = processor(
385
+ text=texts,
386
+ images=image_inputs,
387
+ videos=video_inputs,
388
+ padding=True,
389
+ return_tensors="pt",
390
+ )
391
+ inputs = inputs.to("cuda")
392
+
393
+ # Batch Inference
394
+ generated_ids = model.generate(**inputs, max_new_tokens=128)
395
+ generated_ids_trimmed = [
396
+ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
397
+ ]
398
+ output_texts = processor.batch_decode(
399
+ generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
400
+ )
401
+ print(output_texts)
402
+ ```
403
+ </details>
404
+
405
+ ### More Usage Tips
406
+
407
+ For input images, we support local files, base64, and URLs. For videos, we currently only support local files.
408
+
409
+ ```python
410
+ # You can directly insert a local file path, a URL, or a base64-encoded image into the position where you want in the text.
411
+ ## Local file path
412
+ messages = [
413
+ {
414
+ "role": "user",
415
+ "content": [
416
+ {"type": "image", "image": "file:///path/to/your/image.jpg"},
417
+ {"type": "text", "text": "Describe this image."},
418
+ ],
419
+ }
420
+ ]
421
+ ## Image URL
422
+ messages = [
423
+ {
424
+ "role": "user",
425
+ "content": [
426
+ {"type": "image", "image": "http://path/to/your/image.jpg"},
427
+ {"type": "text", "text": "Describe this image."},
428
+ ],
429
+ }
430
+ ]
431
+ ## Base64 encoded image
432
+ messages = [
433
+ {
434
+ "role": "user",
435
+ "content": [
436
+ {"type": "image", "image": "data:image;base64,/9j/..."},
437
+ {"type": "text", "text": "Describe this image."},
438
+ ],
439
+ }
440
+ ]
441
+ ```
442
+ #### Image Resolution for performance boost
443
+
444
+ The model supports a wide range of resolution inputs. By default, it uses the native resolution for input, but higher resolutions can enhance performance at the cost of more computation. Users can set the minimum and maximum number of pixels to achieve an optimal configuration for their needs, such as a token count range of 256-1280, to balance speed and memory usage.
445
+
446
+ ```python
447
+ min_pixels = 256 * 28 * 28
448
+ max_pixels = 1280 * 28 * 28
449
+ processor = AutoProcessor.from_pretrained(
450
+ "Qwen/Qwen2-VL-2B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels
451
+ )
452
+ ```
453
+
454
+ Besides, We provide two methods for fine-grained control over the image size input to the model:
455
+
456
+ 1. Define min_pixels and max_pixels: Images will be resized to maintain their aspect ratio within the range of min_pixels and max_pixels.
457
+
458
+ 2. Specify exact dimensions: Directly set `resized_height` and `resized_width`. These values will be rounded to the nearest multiple of 28.
459
+
460
+ ```python
461
+ # min_pixels and max_pixels
462
+ messages = [
463
+ {
464
+ "role": "user",
465
+ "content": [
466
+ {
467
+ "type": "image",
468
+ "image": "file:///path/to/your/image.jpg",
469
+ "resized_height": 280,
470
+ "resized_width": 420,
471
+ },
472
+ {"type": "text", "text": "Describe this image."},
473
+ ],
474
+ }
475
+ ]
476
+ # resized_height and resized_width
477
+ messages = [
478
+ {
479
+ "role": "user",
480
+ "content": [
481
+ {
482
+ "type": "image",
483
+ "image": "file:///path/to/your/image.jpg",
484
+ "min_pixels": 50176,
485
+ "max_pixels": 50176,
486
+ },
487
+ {"type": "text", "text": "Describe this image."},
488
+ ],
489
+ }
490
+ ]
491
+ ```
492
+
493
+ ## Limitations
494
+
495
+ While Qwen2-VL are applicable to a wide range of visual tasks, it is equally important to understand its limitations. Here are some known restrictions:
496
+
497
+ 1. Lack of Audio Support: The current model does **not comprehend audio information** within videos.
498
+ 2. Data timeliness: Our image dataset is **updated until June 2023**, and information subsequent to this date may not be covered.
499
+ 3. Constraints in Individuals and Intellectual Property (IP): The model's capacity to recognize specific individuals or IPs is limited, potentially failing to comprehensively cover all well-known personalities or brands.
500
+ 4. Limited Capacity for Complex Instruction: When faced with intricate multi-step instructions, the model's understanding and execution capabilities require enhancement.
501
+ 5. Insufficient Counting Accuracy: Particularly in complex scenes, the accuracy of object counting is not high, necessitating further improvements.
502
+ 6. Weak Spatial Reasoning Skills: Especially in 3D spaces, the model's inference of object positional relationships is inadequate, making it difficult to precisely judge the relative positions of objects.
503
+
504
+ These limitations serve as ongoing directions for model optimization and improvement, and we are committed to continually enhancing the model's performance and scope of application.
505
+
506
+
507
+ ## Citation
508
+
509
+ If you find our work helpful, feel free to give us a cite.
510
+
511
+ ```
512
+ @article{Qwen2VL,
513
+ title={Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution},
514
+ author={Wang, Peng and Bai, Shuai and Tan, Sinan and Wang, Shijie and Fan, Zhihao and Bai, Jinze and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Fan, Yang and Dang, Kai and Du, Mengfei and Ren, Xuancheng and Men, Rui and Liu, Dayiheng and Zhou, Chang and Zhou, Jingren and Lin, Junyang},
515
+ journal={arXiv preprint arXiv:2409.12191},
516
+ year={2024}
517
+ }
518
+
519
+ @article{Qwen-VL,
520
+ title={Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond},
521
+ 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},
522
+ journal={arXiv preprint arXiv:2308.12966},
523
+ year={2023}
524
+ }
525
+ ```