Multi-Style Video Captioner
Multi-Style Video Captioner is an agent that turns any short clip into on brand captions across four requested styles: formal, sarcastic, humorous_tech, and humorous_non_tech. The pipeline samples representative frames across the whole clip with ffmpeg, spacing them by duration, downscaling for speed, and staying under the API payload limits. It then sends those frames to a vision language model (Qwen3 Plus via Fireworks AI) with a carefully engineered, per style system prompt and style specific sampling temperatures for the right balance of accuracy and tone. Every part is configurable: the model, the frame sampling, the per style temperatures, and an optional multi model ensemble with an LLM judge that merges the best candidate captions. A global rate throttle keeps it within API limits, and the whole project ships as a single linux/amd64 Docker image that reads /input/tasks.json and writes valid /output/results.json, mirroring the accuracy and style match evaluation rubric.
Category tags:
For Innovators & Creators