ShotVL

This repository contains ShotVL-3B, a fine-tuned version of Qwen/Qwen2.5-VL-3B-Instruct, developed for expert-level cinematic understanding.

• Paper: ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models
• Project Page: https://vchitect.github.io/ShotBench-project/
• Code: 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, 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, a comprehensive benchmark for evaluating cinematography understanding in vision-language models.

Demo

Image

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 "
    "  and   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

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

  Abbreviations: 
  SS = Shot Size, 
  SF = Shot Framing, 
  CA = Camera Angle, 
  LS = Lens Size, 
  LT = Lighting Type, 
  LC = Lighting Conditions,&n