bluelike commited on
Commit
a28fa9c
1 Parent(s): 1877faf

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +508 -0
README.md ADDED
@@ -0,0 +1,508 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ pipeline_tag: text-generation
6
+ tags:
7
+ - multimodal
8
+ base_model: Qwen/Qwen2-VL-2B-Instruct-AWQ
9
+ ---
10
+
11
+ # Qwen2-VL-2B-Instruct-AWQ
12
+
13
+ ## Introduction
14
+
15
+ We're excited to unveil **Qwen2-VL**, the latest iteration of our Qwen-VL model, representing nearly a year of innovation.
16
+
17
+ ### What’s New in Qwen2-VL?
18
+
19
+ #### Key Enhancements:
20
+
21
+ * **Enhanced Image Comprehension**: We've significantly improved the model's ability to understand and interpret visual information, setting new benchmarks across key performance metrics.
22
+
23
+ * **Advanced Video Understanding**: Qwen2-VL now features superior online streaming capabilities, enabling real-time analysis of dynamic video content with remarkable accuracy.
24
+
25
+ * **Integrated Visual Agent Functionality**: Our model now seamlessly incorporates sophisticated system integration, transforming Qwen2-VL into a powerful visual agent capable of complex reasoning and decision-making.
26
+
27
+ * **Expanded Multilingual Support**: We've broadened our language capabilities to better serve a diverse global user base, making Qwen2-VL more accessible and effective across different linguistic contexts.
28
+
29
+ #### Model Architecture Updates:
30
+
31
+ * **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.
32
+
33
+ * **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.
34
+
35
+ ![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-VL/qwen2_vl.jpg)
36
+
37
+ We have three models with 2, 7 and 72 billion parameters. This repo contains the quantized version of 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).
38
+
39
+
40
+
41
+ ### Benchmark
42
+ #### Performance of Quantized Models
43
+ This section reports the generation performance of quantized models (including GPTQ and AWQ) of the Qwen2-VL series. Specifically, we report:
44
+
45
+ - MMMU_VAL (Accuracy)
46
+ - DocVQA_VAL (Accuracy)
47
+ - MMBench_DEV_EN (Accuracy)
48
+ - MathVista_MINI (Accuracy)
49
+
50
+ We use [VLMEvalkit](https://github.com/kq-chen/VLMEvalKit/tree/add_qwen2vl) to evaluate all models.
51
+
52
+ | Model Size | Quantization | MMMU | DocVQA | MMBench | MathVista |
53
+ | --- | --- | --- | --- | --- | --- |
54
+ Qwen2-VL-2B-Instruct | BF16<br><sup>([🤗](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct)[🤖](https://modelscope.cn/models/qwen/Qwen2-VL-2B-Instruct)) | 41.88 | 88.34 | 72.07 | 44.40 |
55
+ | | GPTQ-Int8<br><sup>([🤗](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct-GPTQ-Int8)[🤖](https://modelscope.cn/models/qwen/Qwen2-VL-2B-Instruct-GPTQ-Int8)) | 41.55 | 88.28 | 71.99 | 44.60 |
56
+ | | GPTQ-Int4<br><sup>([🤗](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct-GPTQ-Int4)[🤖](https://modelscope.cn/models/qwen/Qwen2-VL-2B-Instruct-GPTQ-Int4)) | 39.22 | 87.21 | 70.87 | 41.69 |
57
+ | | AWQ<br><sup>([🤗](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct-AWQ)[🤖](https://modelscope.cn/models/qwen/Qwen2-VL-2B-Instruct-AWQ)) | 41.33 | 86.96 | 71.64 | 39.90 |
58
+
59
+
60
+ #### Speed Benchmark
61
+ This section reports the speed performance of bf16 models, quantized models (including GPTQ-Int4, GPTQ-Int8 and AWQ) of the Qwen2-VL series. Specifically, we report the inference speed (tokens/s) as well as memory footprint (GB) under the conditions of different context lengths.
62
+
63
+ The environment of the evaluation with huggingface transformers is:
64
+
65
+ - NVIDIA A100 80GB
66
+ - CUDA 11.8
67
+ - Pytorch 2.2.1+cu118
68
+ - Flash Attention 2.6.1
69
+ - Transformers 4.38.2
70
+ - AutoGPTQ 0.6.0+cu118
71
+ - AutoAWQ 0.2.5+cu118 (autoawq_kernels 0.0.6+cu118)
72
+
73
+ Note:
74
+
75
+ - We use the batch size of 1 and the least number of GPUs as possible for the evalution.
76
+ - We test the speed and memory of generating 2048 tokens with the input lengths of 1, 6144, 14336, 30720, 63488, and 129024 tokens.
77
+ - 2B (transformers)
78
+
79
+ | Model | Input Length | Quantization | GPU Num | Speed(tokens/s) | GPU Memory(GB) |
80
+ | --- | --- | --- | --- | --- | --- |
81
+ | Qwen2-VL-2B-Instruct | 1 | BF16 | 1 | 35.29 | 4.68 |
82
+ | | | GPTQ-Int8 | 1 | 28.59 | 3.55 |
83
+ | | | GPTQ-Int4 | 1 | 39.76 | 2.91 |
84
+ | | | AWQ | 1 | 29.89 | 2.88 |
85
+ | | 6144 | BF16 | 1 | 36.58 | 10.01 |
86
+ | | | GPTQ-Int8 | 1 | 29.53 | 8.87 |
87
+ | | | GPTQ-Int4 | 1 | 39.27 | 8.21 |
88
+ | | | AWQ | 1 | 33.42 | 8.18 |
89
+ | | 14336 | BF16 | 1 | 36.31 | 17.20 |
90
+ | | | GPTQ-Int8 | 1 | 31.03 | 16.07 |
91
+ | | | GPTQ-Int4 | 1 | 39.89 | 15.40 |
92
+ | | | AWQ | 1 | 32.28 | 15.40 |
93
+ | | 30720 | BF16 | 1 | 32.53 | 31.64 |
94
+ | | | GPTQ-Int8 | 1 | 27.76 | 30.51 |
95
+ | | | GPTQ-Int4 | 1 | 30.73 | 29.84 |
96
+ | | | AWQ | 1 | 31.55 | 29.84 |
97
+
98
+ ## Requirements
99
+ 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:
100
+ ```
101
+ KeyError: 'qwen2_vl'
102
+ ```
103
+
104
+ ## Quickstart
105
+ We offer a toolkit to help you handle various types of visual input more conveniently, as if you were using an API. This includes base64, URLs, and interleaved images and videos. You can install it using the following command:
106
+
107
+ ```bash
108
+ pip install qwen-vl-utils
109
+ ```
110
+
111
+ Here we show a code snippet to show you how to use the chat model with `transformers` and `qwen_vl_utils`:
112
+
113
+ ```python
114
+ from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
115
+ from qwen_vl_utils import process_vision_info
116
+
117
+ # default: Load the model on the available device(s)
118
+ model = Qwen2VLForConditionalGeneration.from_pretrained(
119
+ "Qwen/Qwen2-VL-2B-Instruct-AWQ", torch_dtype="auto", device_map="auto"
120
+ )
121
+
122
+ # We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
123
+ # model = Qwen2VLForConditionalGeneration.from_pretrained(
124
+ # "Qwen/Qwen2-VL-2B-Instruct-AWQ",
125
+ # torch_dtype=torch.bfloat16,
126
+ # attn_implementation="flash_attention_2",
127
+ # device_map="auto",
128
+ # )
129
+
130
+ # default processer
131
+ processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct-AWQ")
132
+
133
+ # 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.
134
+ # min_pixels = 256*28*28
135
+ # max_pixels = 1280*28*28
136
+ # processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct-AWQ", min_pixels=min_pixels, max_pixels=max_pixels)
137
+
138
+ messages = [
139
+ {
140
+ "role": "user",
141
+ "content": [
142
+ {
143
+ "type": "image",
144
+ "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
145
+ },
146
+ {"type": "text", "text": "Describe this image."},
147
+ ],
148
+ }
149
+ ]
150
+
151
+ # Preparation for inference
152
+ text = processor.apply_chat_template(
153
+ messages, tokenize=False, add_generation_prompt=True
154
+ )
155
+ image_inputs, video_inputs = process_vision_info(messages)
156
+ inputs = processor(
157
+ text=[text],
158
+ images=image_inputs,
159
+ videos=video_inputs,
160
+ padding=True,
161
+ return_tensors="pt",
162
+ )
163
+ inputs = inputs.to("cuda")
164
+
165
+ # Inference: Generation of the output
166
+ generated_ids = model.generate(**inputs, max_new_tokens=128)
167
+ generated_ids_trimmed = [
168
+ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
169
+ ]
170
+ output_text = processor.batch_decode(
171
+ generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
172
+ )
173
+ print(output_text)
174
+ ```
175
+ <details>
176
+ <summary>Without qwen_vl_utils</summary>
177
+
178
+ ```python
179
+ from PIL import Image
180
+ import requests
181
+ import torch
182
+ from torchvision import io
183
+ from typing import Dict
184
+ from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
185
+
186
+ # Load the model in half-precision on the available device(s)
187
+ model = Qwen2VLForConditionalGeneration.from_pretrained(
188
+ "Qwen/Qwen2-VL-2B-Instruct-AWQ", device_map="auto"
189
+ )
190
+ processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct-AWQ")
191
+
192
+ # Image
193
+ url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
194
+ image = Image.open(requests.get(url, stream=True).raw)
195
+
196
+ conversation = [
197
+ {
198
+ "role": "user",
199
+ "content": [
200
+ {
201
+ "type": "image",
202
+ },
203
+ {"type": "text", "text": "Describe this image."},
204
+ ],
205
+ }
206
+ ]
207
+
208
+
209
+ # Preprocess the inputs
210
+ text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
211
+ # 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'
212
+
213
+ inputs = processor(
214
+ text=[text_prompt], images=[image], padding=True, return_tensors="pt"
215
+ )
216
+ inputs = inputs.to("cuda")
217
+
218
+ # Inference: Generation of the output
219
+ output_ids = model.generate(**inputs, max_new_tokens=128)
220
+ generated_ids = [
221
+ output_ids[len(input_ids) :]
222
+ for input_ids, output_ids in zip(inputs.input_ids, output_ids)
223
+ ]
224
+ output_text = processor.batch_decode(
225
+ generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
226
+ )
227
+ print(output_text)
228
+ ```
229
+ </details>
230
+
231
+ <details>
232
+ <summary>Multi image inference</summary>
233
+
234
+ ```python
235
+ # Messages containing multiple images and a text query
236
+ messages = [
237
+ {
238
+ "role": "user",
239
+ "content": [
240
+ {"type": "image", "image": "file:///path/to/image1.jpg"},
241
+ {"type": "image", "image": "file:///path/to/image2.jpg"},
242
+ {"type": "text", "text": "Identify the similarities between these images."},
243
+ ],
244
+ }
245
+ ]
246
+
247
+ # Preparation for inference
248
+ text = processor.apply_chat_template(
249
+ messages, tokenize=False, add_generation_prompt=True
250
+ )
251
+ image_inputs, video_inputs = process_vision_info(messages)
252
+ inputs = processor(
253
+ text=[text],
254
+ images=image_inputs,
255
+ videos=video_inputs,
256
+ padding=True,
257
+ return_tensors="pt",
258
+ )
259
+ inputs = inputs.to("cuda")
260
+
261
+ # Inference
262
+ generated_ids = model.generate(**inputs, max_new_tokens=128)
263
+ generated_ids_trimmed = [
264
+ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
265
+ ]
266
+ output_text = processor.batch_decode(
267
+ generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
268
+ )
269
+ print(output_text)
270
+ ```
271
+ </details>
272
+
273
+ <details>
274
+ <summary>Video inference</summary>
275
+
276
+ ```python
277
+ # Messages containing a images list as a video and a text query
278
+ messages = [
279
+ {
280
+ "role": "user",
281
+ "content": [
282
+ {
283
+ "type": "video",
284
+ "video": [
285
+ "file:///path/to/frame1.jpg",
286
+ "file:///path/to/frame2.jpg",
287
+ "file:///path/to/frame3.jpg",
288
+ "file:///path/to/frame4.jpg",
289
+ ],
290
+ "fps": 1.0,
291
+ },
292
+ {"type": "text", "text": "Describe this video."},
293
+ ],
294
+ }
295
+ ]
296
+ # Messages containing a video and a text query
297
+ messages = [
298
+ {
299
+ "role": "user",
300
+ "content": [
301
+ {
302
+ "type": "video",
303
+ "video": "file:///path/to/video1.mp4",
304
+ "max_pixels": 360 * 420,
305
+ "fps": 1.0,
306
+ },
307
+ {"type": "text", "text": "Describe this video."},
308
+ ],
309
+ }
310
+ ]
311
+
312
+ # Preparation for inference
313
+ text = processor.apply_chat_template(
314
+ messages, tokenize=False, add_generation_prompt=True
315
+ )
316
+ image_inputs, video_inputs = process_vision_info(messages)
317
+ inputs = processor(
318
+ text=[text],
319
+ images=image_inputs,
320
+ videos=video_inputs,
321
+ padding=True,
322
+ return_tensors="pt",
323
+ )
324
+ inputs = inputs.to("cuda")
325
+
326
+ # Inference
327
+ generated_ids = model.generate(**inputs, max_new_tokens=128)
328
+ generated_ids_trimmed = [
329
+ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
330
+ ]
331
+ output_text = processor.batch_decode(
332
+ generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
333
+ )
334
+ print(output_text)
335
+ ```
336
+ </details>
337
+
338
+ <details>
339
+ <summary>Batch inference</summary>
340
+
341
+ ```python
342
+
343
+ # Sample messages for batch inference
344
+ messages1 = [
345
+ {
346
+ "role": "user",
347
+ "content": [
348
+ {"type": "image", "image": "file:///path/to/image1.jpg"},
349
+ {"type": "image", "image": "file:///path/to/image2.jpg"},
350
+ {"type": "text", "text": "What are the common elements in these pictures?"},
351
+ ],
352
+ }
353
+ ]
354
+ messages2 = [
355
+ {"role": "system", "content": "You are a helpful assistant."},
356
+ {"role": "user", "content": "Who are you?"},
357
+ ]
358
+ # Combine messages for batch processing
359
+ messages = [messages1, messages1]
360
+
361
+ # Preparation for batch inference
362
+ texts = [
363
+ processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True)
364
+ for msg in messages
365
+ ]
366
+ image_inputs, video_inputs = process_vision_info(messages)
367
+ inputs = processor(
368
+ text=texts,
369
+ images=image_inputs,
370
+ videos=video_inputs,
371
+ padding=True,
372
+ return_tensors="pt",
373
+ )
374
+ inputs = inputs.to("cuda")
375
+
376
+ # Batch Inference
377
+ generated_ids = model.generate(**inputs, max_new_tokens=128)
378
+ generated_ids_trimmed = [
379
+ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
380
+ ]
381
+ output_texts = processor.batch_decode(
382
+ generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
383
+ )
384
+ print(output_texts)
385
+ ```
386
+ </details>
387
+
388
+ ### More Usage Tips
389
+
390
+ For input images, we support local files, base64, and URLs. For videos, we currently only support local files.
391
+
392
+ ```python
393
+ # You can directly insert a local file path, a URL, or a base64-encoded image into the position where you want in the text.
394
+ ## Local file path
395
+ messages = [
396
+ {
397
+ "role": "user",
398
+ "content": [
399
+ {"type": "image", "image": "file:///path/to/your/image.jpg"},
400
+ {"type": "text", "text": "Describe this image."},
401
+ ],
402
+ }
403
+ ]
404
+ ## Image URL
405
+ messages = [
406
+ {
407
+ "role": "user",
408
+ "content": [
409
+ {"type": "image", "image": "http://path/to/your/image.jpg"},
410
+ {"type": "text", "text": "Describe this image."},
411
+ ],
412
+ }
413
+ ]
414
+ ## Base64 encoded image
415
+ messages = [
416
+ {
417
+ "role": "user",
418
+ "content": [
419
+ {"type": "image", "image": "data:image;base64,/9j/..."},
420
+ {"type": "text", "text": "Describe this image."},
421
+ ],
422
+ }
423
+ ]
424
+
425
+ ```
426
+ #### Image Resolution for performance boost
427
+
428
+ 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.
429
+
430
+ ```python
431
+ min_pixels = 256 * 28 * 28
432
+ max_pixels = 1280 * 28 * 28
433
+ processor = AutoProcessor.from_pretrained(
434
+ "Qwen/Qwen2-VL-2B-Instruct-AWQ", min_pixels=min_pixels, max_pixels=max_pixels
435
+ )
436
+ ```
437
+
438
+ Besides, We provide two methods for fine-grained control over the image size input to the model:
439
+
440
+ 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.
441
+
442
+ 2. Specify exact dimensions: Directly set `resized_height` and `resized_width`. These values will be rounded to the nearest multiple of 28.
443
+
444
+ ```python
445
+ # min_pixels and max_pixels
446
+ messages = [
447
+ {
448
+ "role": "user",
449
+ "content": [
450
+ {
451
+ "type": "image",
452
+ "image": "file:///path/to/your/image.jpg",
453
+ "resized_height": 280,
454
+ "resized_width": 420,
455
+ },
456
+ {"type": "text", "text": "Describe this image."},
457
+ ],
458
+ }
459
+ ]
460
+ # resized_height and resized_width
461
+ messages = [
462
+ {
463
+ "role": "user",
464
+ "content": [
465
+ {
466
+ "type": "image",
467
+ "image": "file:///path/to/your/image.jpg",
468
+ "min_pixels": 50176,
469
+ "max_pixels": 50176,
470
+ },
471
+ {"type": "text", "text": "Describe this image."},
472
+ ],
473
+ }
474
+ ]
475
+ ```
476
+
477
+ ## Limitations
478
+
479
+ 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:
480
+
481
+ 1. Lack of Audio Support: The current model does **not comprehend audio information** within videos.
482
+ 2. Data timeliness: Our image dataset is **updated until June 2023**, and information subsequent to this date may not be covered.
483
+ 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.
484
+ 4. Limited Capacity for Complex Instruction: When faced with intricate multi-step instructions, the model's understanding and execution capabilities require enhancement.
485
+ 5. Insufficient Counting Accuracy: Particularly in complex scenes, the accuracy of object counting is not high, necessitating further improvements.
486
+ 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.
487
+
488
+ 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.
489
+
490
+
491
+ ## Citation
492
+
493
+ If you find our work helpful, feel free to give us a cite.
494
+
495
+ ```
496
+ @article{Qwen2-VL,
497
+ title={Qwen2-VL},
498
+ author={Qwen team},
499
+ year={2024}
500
+ }
501
+
502
+ @article{Qwen-VL,
503
+ title={Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond},
504
+ 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},
505
+ journal={arXiv preprint arXiv:2308.12966},
506
+ year={2023}
507
+ }
508
+ ```