This code implements an automated video matching and editing functionality. It can intelligently select relevant video clips from a video library based on the given scene descriptions, and generate a complete video work.
json.py first extracts the video addresses from the JSON file, and retrieves the corresponding video information, including resolution, duration, etc., and saves them as a new JSON file with more complete information, organized by category.
dl.py downloads the corresponding videos based on the JSON file. The filenames include a lot of video-related information.
caption_glm-4.py uses the first frame of the video and a rough caption as input, and calls the multi-modal large model https://github.com/THUDM/GLM-4 to complete the precise understanding of the video.
Finally, demo1.py and demo2.py recall and match the video clips based on the keyword input, and then assemble them into a complete video.
To run the demo, execute the following command in your terminal:
python demo2.py