SVLM Utils

SVLM Utils - Video and Image Processing Utilities for SVECTOR Video Language Model (SVLM)

Installation

pip install svlm-utils

Quick Start

from svlm_utils import process_vision_info, fetch_image, fetch_video

# Process messages with images/videos
messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Describe this video"},
            {"type": "video", "video": "path/to/video.mp4", "max_fps": 8},
        ],
    }
]

images, videos, video_kwargs = process_vision_info(messages)

Features

• 🎬 Video Processing: Load and sample video frames with configurable FPS
• 🖼️ Image Processing: Fetch images from URLs, base64, or local files
• 🔧 Multiple Backends: Support for decord, torchcodec, and PyAV
• ⚡ Optimized Sampling: Intelligent frame sampling for efficient processing

Video Loading Options

# Load video with specific settings
video_config = {
    "video": "path/to/video.mp4",
    "max_fps": 8,            # Maximum frames per second to sample
    "num_frames": 64,        # Maximum total frames
    "clip": (0, 10),         # Optional: extract specific time range (seconds)
}

video_frames, metadata = fetch_video(video_config)

Supported Formats

Images

• Local files: image.jpg, image.png
• URLs: http://..., https://...
• Base64: data:image/jpeg;base64,...
• PIL Images: Direct PIL.Image objects

Videos

• Local files: .mp4, .avi, .mov, .webm
• File URLs: file:///path/to/video.mp4

License

Apache License 2.0

Author

SVECTOR Corporation - https://svector.co