This repo at a glanceFrom the README
Copy in, finished video out. A local-first open-source short-video production line: give it a topic, and it writes the script, finds visuals, voices the narration, burns in subtitles, and outputs the finished video and publishing copy. By default you can run the first video at zero cost, and it tells you before running how much it will cost and how long it will take. The visuals are not slideshows: for each shot it chooses between real-footage video clips and photos (if a vision model is available it scores and selects, otherwise it follows retrieval order). On a real machine, a 30-second video has 5 shots, 3 of which are real-footage clips; the two photo shots also move. The whole image is centered without cropping the subject, with a blurred, enlarged copy of itself behind it that slowly shifts, alternating direction from shot to shot. If the asset library has no match, it renders an image locally (through the same rendering path, with the same camera movement). The AI short-drama line uses genuine AI-generated video (locally with MiniMax-H3 or in the cloud).