StyleFrame Caption Agent

StyleFrame Caption Agent is a containerized AI agent built for the Video Captioning track. It reads video-captioning tasks from /input/tasks.json, downloads each video clip, samples representative frames with FFmpeg, compresses the visual timeline into a contact sheet, and uses a Fireworks multimodal model to generate grounded visual facts and four caption styles in a single pass. The system is designed for hidden-test reliability rather than only fitting the public examples. It supports formal, sarcastic, humorous-tech, and humorous-non-tech captions, while applying local guardrails for length, tone, tech-vs-non-tech separation, scene consistency, and common hallucination risks. If the single-pass model output fails to parse, the agent can fall back to a two-stage pipeline that first extracts visual facts and then rewrites captions. The project also includes local benchmarking and judging tools to compare configurations by runtime, output validity, caption quality, and style match. The Docker image is linux/amd64 compatible and writes the required /output/results.json file for evaluation.