File size: 10,435 Bytes
d6223df
 
 
12a1a89
babd410
 
 
 
12a1a89
 
2553d0c
12a1a89
2553d0c
 
fd8c356
12a1a89
 
9143ef9
1724675
12a1a89
1724675
2553d0c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9659ded
2553d0c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12a1a89
 
 
 
 
 
 
 
 
 
2553d0c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12a1a89
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ca8673
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
---
base_model:
- Qwen/Qwen2.5-VL-7B-Instruct
license: apache-2.0
tags:
- vision-language
- cinematography
- shotbench
pipeline_tag: image-text-to-text
library_name: transformers
---

## Model description

This model is a fine-tuned version of [Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), introduced in the paper [ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models](https://huggingface.co/papers/2506.21356). It is trained on the largest and high-quality dataset for cinematic language understanding to date. It currently achieves state-of-the-art performance on [ShotBench](https://vchitect.github.io/ShotBench-project/), a comprehensive benchmark for evaluating cinematography understanding in vision-language models.

**Project Page:** [https://vchitect.github.io/ShotBench-project/](https://vchitect.github.io/ShotBench-project/)

**Code:** [https://github.com/Vchitect/ShotBench](https://github.com/Vchitect/ShotBench)

### Demo

**Image**

```python
import cv2
import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

device = "cuda"
device_map = "balanced"
dtype = torch.bfloat16
image_path = "/path/to/image.jpg"

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
  "Vchitect/ShotVL-7B",
  device_map=device_map,
  attn_implementation="flash_attention_2",
  torch_dtype=dtype,
).eval()
processor = AutoProcessor.from_pretrained(
  "Vchitect/ShotVL-7B", revision="refs/pr/24", use_fast=True, torch_dtype=dtype
)

msgs = [
  {"role": "system", "content": "You are a helpful assistant."},
  {
    "role": "user",
    "content": [
      {"type": "image", "image": image_path},
      {"type": "text", "text": "What's the shot size of this shot?"},
    ],
  },
]

text = processor.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(msgs)
inputs = processor(
  text=[text],
  images=image_inputs,
  videos=video_inputs,
  padding=True,
  return_tensors="pt",
).to(device)

with torch.inference_mode():
  out_ids = model.generate(**inputs, max_new_tokens=640)
  
trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, out_ids)]
print(processor.batch_decode(trimmed, skip_special_tokens=True)[0])
```

**Video**

```python
import cv2
import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

device = "cuda"
device_map = "balanced"
dtype = torch.bfloat16
video_path = "/path/to/video.mp4"

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
  "Vchitect/ShotVL-7B",
  device_map=device_map,
  attn_implementation="flash_attention_2",
  torch_dtype=dtype,
).eval()
processor = AutoProcessor.from_pretrained(
  "Vchitect/ShotVL-7B", revision="refs/pr/24", use_fast=True, torch_dtype=dtype
)

question = (
    "What's the camera movement in this movie shot?
"
    "Options:
A. Boom down
B. Boom up
C. Push in
D. Pull out
"
    "Please select the most likely answer from the options above.
"
)
msgs = [
  {"role": "system", "content": "You are a helpful assistant."},
  {
    "role": "user",
    "content": [
      {"type": "video", "video": video_path, "max_pixels": 360*640, "fps": 12.0},
      {"type": "text", "text": question},
    ],
  },
]

text = processor.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(msgs)
inputs = processor(
  text=[text],
  images=image_inputs,
  videos=video_inputs,
  padding=True,
  return_tensors="pt",
).to(device)

with torch.inference_mode():
  out_ids = model.generate(**inputs, max_new_tokens=640)
  
trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, out_ids)]
print(processor.batch_decode(trimmed, skip_special_tokens=True)[0])
```

## Evaluation Results

<div align="center">
<table>
  <caption>
    <small>
      Abbreviations:&nbsp;
      SS = <em>Shot&nbsp;Size</em>,&nbsp;
      SF = <em>Shot&nbsp;Framing</em>,&nbsp;
      CA = <em>Camera&nbsp;Angle</em>,&nbsp;
      LS = <em>Lens&nbsp;Size</em>,&nbsp;
      LT = <em>Lighting&nbsp;Type</em>,&nbsp;
      LC = <em>Lighting&nbsp;Conditions</em>,&nbsp;
      SC = <em>Shot&nbsp;Composition</em>,&nbsp;
      CM = <em>Camera&nbsp;Movement</em>.&nbsp;
      <u>Underline</u> marks previous best in each group.<br>
      <strong>Our <em>ShotVL</em> models establish new SOTA.</strong>
    </small>
  </caption><thead>
    <tr>
      <th>Models</th><th>SS</th><th>SF</th><th>CA</th><th>LS</th><th>LT</th>
      <th>LC</th><th>SC</th><th>CM</th><th>Avg</th>
    </tr>
  </thead><tbody>
  <tr><th colspan="10"><em>Open-Sourced&nbsp;VLMs</em></th></tr>
                            <tr><td>Qwen2.5-VL-3B-Instruct</td><td>54.6</td><td>56.6</td><td>43.1</td><td>36.6</td><td>59.3</td><td>45.1</td><td>41.5</td><td>31.9</td><td>46.1</td></tr>
                            <tr><td>Qwen2.5-VL-7B-Instruct</td><td>69.1</td><td>73.5</td><td>53.2</td><td>47.0</td><td>60.5</td><td>47.4</td><td>49.9</td><td>30.2</td><td>53.8</td></tr>
                            <tr><td>LLaVA-NeXT-Video-7B</td><td>35.9</td><td>37.1</td><td>32.5</td><td>27.8</td><td>50.9</td><td>31.7</td><td>28.0</td><td>31.3</td><td>34.4</td></tr>
                            <tr><td>LLaVA-Video-7B-Qwen2</td><td>56.9</td><td>65.4</td><td>45.1</td><td>36.0</td><td>63.5</td><td>45.4</td><td>37.4</td><td>35.3</td><td>48.1</td></tr>
                            <tr><td>LLaVA-Onevision-Qwen2-7B-Ov-Chat</td><td>58.4</td><td>71.0</td><td>52.3</td><td>38.7</td><td>59.5</td><td>44.9</td><td>50.9</td><td>39.7</td><td>51.9</td></tr>
                            <tr><td>InternVL2.5-8B</td><td>56.3</td><td>70.3</td><td>50.8</td><td>41.1</td><td>60.2</td><td>45.1</td><td>50.1</td><td>33.6</td><td>50.9</td></tr>
                            <tr><td>InternVL3-2B</td><td>56.3</td><td>56.0</td><td>44.4</td><td>34.6</td><td>56.8</td><td>44.6</td><td>43.0</td><td>38.1</td><td>46.7</td></tr>
                            <tr><td>InternVL3-8B</td><td>62.1</td><td>65.8</td><td>46.8</td><td>42.9</td><td>58.0</td><td>44.3</td><td>46.8</td><td>44.2</td><td>51.4</td></tr>
                            <tr><td>InternVL3-14B</td><td>59.6</td><td>82.2</td><td>55.4</td><td>40.7</td><td>61.7</td><td>44.6</td><td>51.1</td><td>38.2</td><td>54.2</td></tr>
                            <tr><td>Internlm-xcomposer2d5-7B</td><td>51.1</td><td>71.0</td><td>39.8</td><td>32.7</td><td>59.3</td><td>35.7</td><td>35.7</td><td>38.8</td><td>45.5</td></tr>
                            <tr><td>Ovis2-8B</td><td>35.9</td><td>37.1</td><td>32.5</td><td>27.8</td><td>50.9</td><td>31.7</td><td>28.0</td><td>35.3</td><td>34.9</td></tr>
                            <tr><td>VILA1.5-3B</td><td>33.4</td><td>44.9</td><td>32.1</td><td>28.6</td><td>50.6</td><td>35.7</td><td>28.4</td><td>21.5</td><td>34.4</td></tr>
                            <tr><td>VILA1.5-8B</td><td>40.6</td><td>44.5</td><td>39.1</td><td>29.7</td><td>48.9</td><td>32.9</td><td>34.4</td><td>36.9</td><td>38.4</td></tr>
                            <tr><td>VILA1.5-13B</td><td>36.7</td><td>54.6</td><td>40.7</td><td>34.8</td><td>52.8</td><td>35.4</td><td>34.2</td><td>31.3</td><td>40.1</td></tr>
                            <tr><td>Instructblip-vicuna-7B</td><td>27.0</td><td>27.9</td><td>34.5</td><td>29.4</td><td>44.4</td><td>29.7</td><td>27.1</td><td>25.0</td><td>30.6</td></tr>
                            <tr><td>Instructblip-vicuna-13B</td><td>26.8</td><td>29.2</td><td>27.9</td><td>28.0</td><td>39.0</td><td>24.0</td><td>27.1</td><td>22.0</td><td>28.0</td></tr>
                            <tr><td>InternVL2.5-38B</td><td>67.8</td><td><u>85.4</u></td><td>55.4</td><td>41.7</td><td>61.7</td><td>48.9</td><td>52.4</td><td>44.0</td><td>57.2</td></tr>
                            <tr><td>InternVL3-38B</td><td>68.0</td><td>84.0</td><td>51.9</td><td>43.6</td><td>64.4</td><td>46.9</td><td>54.7</td><td>44.6</td><td>57.3</td></tr>
                            <tr><td>Qwen2.5-VL-32B-Instruct</td><td>62.3</td><td>76.6</td><td>51.0</td><td>48.3</td><td>61.7</td><td>44.0</td><td>52.2</td><td>43.8</td><td>55.0</td></tr>
                            <tr><td>Qwen2.5-VL-72B-Instruct</td><td><u>75.1</u></td><td>82.9</td><td>56.7</td><td>46.8</td><td>59.0</td><td><u>49.4</u></td><td>54.1</td><td><u>48.9</u></td><td>59.1</td></tr>
                            <tr><td>InternVL3-78B</td><td>69.7</td><td>80.0</td><td>54.5</td><td>44.0</td><td><u>65.5</u></td><td>47.4</td><td>51.8</td><td>44.4</td><td>57.2</td></tr>
<tr><th colspan="10"><em>Proprietary&nbsp;VLMs</em></th></tr>
                            <tr><td>Gemini-2.0-flash</td><td>48.9</td><td>75.5</td><td>44.6</td><td>31.9</td><td>62.2</td><td>48.9</td><td>52.4</td><td>47.4</td><td>51.5</td></tr>
                            <tr><td>Gemini-2.5-flash-preview-04-17</td><td>57.7</td><td>82.9</td><td>51.4</td><td>43.8</td><td>65.2</td><td>45.7</td><td>45.9</td><td>43.5</td><td>54.5</td></tr>
                            <tr><td>GPT-4o</td><td>69.3</td><td>83.1</td><td><u>58.2</u></td><td><u>48.9</u></td><td>63.2</td><td>48.0</td><td><u>55.2</u></td><td>48.3</td><td><u>59.3</u></td></tr>
<tr><th colspan="10"><em>Ours</em></th></tr>
<tr>
  <td>ShotVL-3B
    <a href="https://huggingface.co/Vchitect/ShotVL-3B">
      <img src="https://img.shields.io/badge/Model-HF-yellow?logo=huggingface" alt="HF">
    </a>
  </td>
  <td>77.9</td><td>85.6</td><td>68.8</td><td>59.3</td><td>65.7</td>
  <td>53.1</td><td>57.4</td><td>51.7</td><td>65.1</td>
</tr>
<tr>
  <td>ShotVL-7B
    <a href="https://huggingface.co/Vchitect/ShotVL-7B">
      <img src="https://img.shields.io/badge/Model-HF-yellow?logo=huggingface" alt="HF">
    </a>
  </td>
  <td>81.2</td><td>90.1</td><td>78.0</td><td>68.5</td><td>70.1</td>
  <td>64.3</td><td>45.7</td><td>62.9</td><td>70.1</td>
</tr>  </tbody>
</table></div>

## BibTeX

```
@misc{
      liu2025shotbench,
      title={ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models}, 
      author={Hongbo Liu and Jingwen He and Yi Jin and Dian Zheng and Yuhao Dong and Fan Zhang and Ziqi Huang and Yinan He and Yangguang Li and Weichao Chen and Yu Qiao and Wanli Ouyang and Shengjie Zhao and Ziwei Liu},
      year={2025},
      eprint={2506.21356},
      achivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.21356}, 
    }
```