README.md

---
datasets:
- yaolily/GenS-Video-150K
license: other # Research use only
---

<p align="center">
🔗 <a href="https://generative-sampler.github.io/" target="_blank">Project Page</a> · 📖 <a href="https://arxiv.org/abs/2503.09146" target="_blank">Paper</a> · ⭐ <a href="https://github.com/yaolinli/GenS" target="_blank">GitHub</a> · 📊 <a href="https://huggingface.co/datasets/yaolily/GenS-Video-150K" target="_blank">Dataset</a> · 🤗 <a href="https://huggingface.co/yaolily/GenS" target="_blank">Checkpoints</a>
</p>

## Model Description

**GenS** (Generative Frame Sampler) is a novel approach that identifies question-relevant frames from long videos spanning minutes to hours. Given a long video and a user question, GenS effectively searches through the original massive collection of frames to produce a concise selection and enhances the performance of downstream VideoQA Assistants (such as Qwen2-VL, LLaVA-Video, VILA-v1.5, and Aria) by providing fewer but more informative frames.

GenS is built upon advanced long-context VideoLLMs (such as Aria and Qwen2.5VL), transforming key frame sampling into a generative task.

<img src="https://generative-sampler.github.io/static/images/teaser.png" alt="GenS Framework" style="width: 100%;">

## Key Features of GenS

✨ **Temporal Understanding:**
GenS effectively captures temporal relationships between successive frames, enabling complex reasoning about temporal sequences such as "immediately after" events in videos.

📝 **Complex Instruction Understanding:**
Powered by built-in LLMs, GenS comprehends complex and flexible textual instructions, allowing it to interpret nuanced queries and identify the most relevant visual content.

⚡ **Effective Video-Text Alignment:**
Its native multi-modal architecture enables sophisticated multi-hop reasoning by seamlessly aligning long-range temporal cues with language semantics, resulting in more accurate frame selection.

🎉 **State-of-the-Art Performance:**
GenS significantly boosts the performance of various VideoQA models, achieving SOTA results on long-form video benchmarks when integrated with open-source models.

## Performance Highlights
-   🏆 **LongVideoBench**: LLaVA-Video-72B w/ GenS achieves **66.8** accuracy (+4.3)
-   🏆 **MLVU**: LLaVA-Video-72B w/ GenS achieves **77.0** accuracy (+2.7)
-   🏆 **HourVideo**: Aria w/ GenS obtains **39.2** accuracy, while Gemini-1.5-pro w/ GenS obtains **40.7** accuracy


<img src="https://generative-sampler.github.io/static/images/table_main.png" alt="Main Results Table" style="width: 100%;">
<img src="https://generative-sampler.github.io/static/images/hourvideo.png" alt="HourVideo Results Table" style="width: 100%;">

## Quick Start

### Installation
After creating your conda environment, install the required dependencies:
```
pip install transformers==4.45.0 accelerate==0.34.1 sentencepiece==0.2.0 torchvision requests torch Pillow
pip install flash-attn --no-build-isolation
```

### Usage

```
import torch
from PIL import Image
import sys
import os
from typing import List

# Import required libraries
from transformers import AutoProcessor, AutoTokenizer, AutoConfig, AutoModel, AutoModelForCausalLM
from yivl.yivl_model_hf import YiVLForConditionalGeneration, YiVLConfig
from yivl.siglip_navit_490 import NaViTProcessor
from yivl.constants import (
    DEFAULT_IMAGE_END_TOKEN,
    DEFAULT_IMAGE_START_TOKEN,
    DEFAULT_IMAGE_TOKEN,
    IMAGE_TOKEN_INDEX,
)
from deepseekv1moe.modeling_deepseek import DeepseekConfig, DeepseekForCausalLM


def setup_model():
    """Set up and load the GenS model and its components."""
    
    # Register custom models with the Auto classes
    AutoConfig.register("yi_vl", YiVLConfig)
    AutoModel.register(YiVLConfig, YiVLForConditionalGeneration)
    AutoConfig.register("deepseek", DeepseekConfig)
    AutoModelForCausalLM.register(DeepseekConfig, DeepseekForCausalLM)
    
    # Load model from Hugging Face
    model_id = "yaolily/GenS"
    
    # Load configuration
    config = AutoConfig.from_pretrained(model_id)
    
    # Load model with optimizations
    model = AutoModel.from_pretrained(
        model_id,
        attn_implementation="flash_attention_2",
        low_cpu_mem_usage=True,
        torch_dtype=torch.bfloat16
    ).to(torch.device("cuda"))
    
    # Load tokenizer with special token handling
    tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True)
    if not tokenizer.pad_token or tokenizer.pad_token_id < 0:
        try:
            tokenizer.add_special_tokens({"pad_token": "<unk>"})
            if tokenizer.pad_token_id is None:
                tokenizer.add_special_tokens({"pad_token": "<mask>"})
        except ValueError:
            tokenizer.add_special_tokens({"pad_token": "<|endoftext|>"})
    
    # Initialize the custom image processor
    processor = NaViTProcessor(image_max_size=490)
    
    print("GenS Model loaded successfully!")
    return model, tokenizer, processor


def gens_frame_sampler(question: str, frame_paths: List[str], model, tokenizer, processor):
    """
    Use GenS model to identify and score relevant frames for a video question.
    
    Args:
        question: The question to answer about the video
        frame_paths: List of paths to video frames
        model: Pre-loaded GenS model
        tokenizer: Pre-loaded tokenizer
        processor: Pre-loaded image processor
        
    Returns:
        The model's response with relevance scores for frames
    """
    # Load frames as PIL images
    frames = []
    for path in frame_paths:
        try:
            img = Image.open(path).convert("RGB")
            # Optional: resize images to expected size
            if img.width > 490 or img.height > 490:
                ratio = min(490/img.width, 490/img.height)
                new_size = (int(img.width * ratio), int(img.height * ratio))
                img = img.resize(new_size)
            frames.append(img)
        except Exception as e:
            print(f"Error loading image {path}: {e}")
    
    if not frames:
        return "Error: No valid frames could be loaded"
    
    # Create prompt
    prompt = """Please identify the video frames most relevant to the given question and provide 
              their timestamps in seconds along with a relevance score. The score should be on a 
              scale from 1 to 5, where higher scores indicate greater relevance. Return the output 
              strictly in the following JSON format: {"timestamp": score, ...}."""
    
    # Format the input as expected by the model
    frm_placeholders = ["<image1>" for _ in range(len(frames))]
    content = "{}Question: {}\n{}".format("".join(frm_placeholders), question, prompt)
    question_data = [{"role": "user", "content": content}]
    
    # Apply chat template
    formatted_question = tokenizer.apply_chat_template(question_data, add_generation_prompt=True, tokenize=False)
    
    # Process the images and text
    inputs = processor(
        text=[formatted_question],
        images=frames,
        padding=True,
        return_tensors="pt"
    )
    inputs = {k: v.to(model.device) for k, v in inputs.items()}
    
    # Generate the response
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=256,
            do_sample=False,
            temperature=0.0
        )
    
    # Decode and extract the relevant part of the response
    response = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    result = response.split("assistant\n")[-1].split("<|im_end|>")[0].strip()
    
    return result


# Example usage
if __name__ == "__main__":
    # Load model components
    model, tokenizer, processor = setup_model()
    
    # Example video frames (replace with your actual paths)
    frame_paths = [
        "/path/to/video/frames/00001.jpg", 
        "/path/to/video/frames/00002.jpg",
        # Add more frames...
    ]
    
    # Example question
    question = "Which frames show a person opening the door?"
    
    # Get frame relevance scores
    result = gens_frame_sampler(question, frame_paths, model, tokenizer, processor)
    
    print(f"Question: {question}")
    print(f"Relevant frames with scores: {result}")
```
**Output Format:**
The model returns relevance scores for frames in JSON format
Example output: `{"15": 5, "16": 4, "45-46": 3, ...}` means frame indexing 15 has relevance score 5, frame indexing 16 has relevance score 4, frame indexing 45-46 has relevance score 3, ...



## Citation
If you find our work helpful, please consider citing.
```
@article{yao2025generative,
    title={Generative Frame Sampler for Long Video Understanding},
    author={Yao, Linli and Wu, Haoning and Ouyang, Kun and Zhang, Yuanxing and Xiong, Caiming and Chen, Bei and Sun, Xu and Li, Junnan},
    journal={arXiv preprint arXiv:2503.09146},
    year={2025}
}
```