AMD Captivate AI

AMD Captivate AI is an end-to-end intelligent video captioning and understanding system built for the AMD Developer Hackathon 2026. The project automates the creation of high-quality descriptive captions tailored to different target audiences. On container startup, the batch processor dynamically reads video URLs from an input file, downloads them, and utilizes computer vision (OpenCV) to perform intelligent keyframe extraction. By applying a grayscale structural difference algorithm, it automatically filters out near-duplicate frames to minimize token usage and focus only on unique, representative visual changes. These keyframes are captioned using the Fireworks AI Qwen 3.5/3.7 Plus Vision model. The individual frame descriptions are consolidated to generate a single, cohesive overall video summary. Finally, the system rewrites this summary into four distinct styles to fit different contexts: Formal (professional and objective), Sarcastic (witty and mocking), Humorous-Tech (loaded with coding and tech references), and Humorous-NonTech (everyday humor). The results are written into a structured JSON file matching the Track 2 specifications. The system also supports a responsive React dashboard for local deployment, featuring SRT subtitle exporting, visual keyframe timelines, and styled captions.