Four Voices, One Clip

AMD Video Captioning turns a single short clip into four captions, each in a distinct voice: formal, sarcastic, humorous-tech, and humorous-non-tech. The guiding idea is "facts first, personality second." Rather than captioning blind, the app uses ffmpeg to sample keyframes evenly across the clip and sends them to a vision-capable model on the Fireworks AI API (kimi-k2p6), which returns a factual description of what actually happens on screen. That grounded understanding is then re-voiced into all four styles, so every caption stays accurate while sounding completely different. A built-in LLM judge scores each caption on accuracy and tone, the same rubric the contest uses, making quality visible before submission. The project ships as a FastAPI backend with a zero-build web UI for selecting styles, a batch runner that produces a submission file for the entire clip set, and a swappable provider that falls back to an offline mock when no API key is present. Fireworks model IDs live in a single config value, so swapping models or dropping in a fine-tuned captioner is a one-line change.