File size: 12,619 Bytes
3ef31d8 c81fc44 3e703ac c81fc44 3ef31d8 c81fc44 a2503a4 c81fc44 3ef31d8 c81fc44 a2503a4 3ef31d8 c4de598 3ef31d8 e1120f0 3ef31d8 e1120f0 3ef31d8 c81fc44 3ef31d8 c81fc44 3ef31d8 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 |
---
base_model:
- Qwen/Qwen2.5-VL-3B-Instruct
license: apache-2.0
tags:
- vision-language
- cinematography
- shotbench
pipeline_tag: image-text-to-text
library_name: transformers
---
# ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models
This repository contains **ShotVL-3B**, a fine-tuned version of [Qwen/Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct), developed for expert-level cinematic understanding.
* **Paper:** [ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models](https://arxiv.org/abs/2506.21356)
* **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)
## Abstract
Cinematography, the fundamental visual language of film, is essential for conveying narrative, emotion, and aesthetic quality. While recent Vision-Language Models (VLMs) demonstrate strong general visual understanding, their proficiency in comprehending the nuanced cinematic grammar embedded within individual shots remains largely unexplored and lacks robust evaluation. This critical gap limits both fine-grained visual comprehension and the precision of AI-assisted video generation. To address this, we introduce ShotBench, a comprehensive benchmark specifically designed for cinematic language understanding. It features over 3.5k expert-annotated QA pairs from images and video clips, meticulously curated from over 200 acclaimed (predominantly Oscar-nominated) films and spanning eight key cinematography dimensions. Our evaluation of 24 leading VLMs on ShotBench reveals their substantial limitations: even the top-performing model achieves less than 60% average accuracy, particularly struggling with fine-grained visual cues and complex spatial reasoning. To catalyze advancement in this domain, we construct ShotQA, a large-scale multimodal dataset comprising approximately 70k cinematic QA pairs. Leveraging ShotQA, we develop ShotVL through supervised fine-tuning and Group Relative Policy Optimization. ShotVL significantly outperforms all existing open-source and proprietary models on ShotBench, establishing new state-of-the-art performance. We open-source our models, data, and code to foster rapid progress in this crucial area of AI-driven cinematic understanding and generation.
## Model description
This model is a fine-tuned version of [Qwen/Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct), trained by supervised fine-tuning and GRPO 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.
### 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-3B",
device_map=device_map,
attn_implementation="flash_attention_2",
torch_dtype=dtype,
).eval()
processor = AutoProcessor.from_pretrained(
"Vchitect/ShotVL-3B", revision="refs/pr/24", use_fast=True, torch_dtype=dtype
)
SYSTEM_PROMPT = (
"A conversation between User and Assistant. The user asks a question, and the Assistant "
"solves it. The assistant first thinks about the reasoning process in the mind and then "
"provides the user with the answer. The reasoning process and answer are enclosed within "
"<think> </think> and <answer> </answer> tags."
)
msgs = [
{"role": "system", "content": SYSTEM_PROMPT},
{
"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-3B",
device_map=device_map,
attn_implementation="flash_attention_2",
torch_dtype=dtype,
).eval()
processor = AutoProcessor.from_pretrained(
"Vchitect/ShotVL-3B", 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": "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:
SS = <em>Shot Size</em>,
SF = <em>Shot Framing</em>,
CA = <em>Camera Angle</em>,
LS = <em>Lens Size</em>,
LT = <em>Lighting Type</em>,
LC = <em>Lighting Conditions</em>,
SC = <em>Shot Composition</em>,
CM = <em>Camera Movement</em>.
<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 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 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},
}
``` |