VividStudio

VividStudio shifts video AI from fragile prompt-wrappers to a resilient, production-grade agentic pipeline designed for both high-stakes grading environments and real-world creator privacy. The Core Pipeline: We structurally decouple observation from performance. Using FFmpeg for scene detection, our Vision Engine outputs a strict Pydantic "Fact-Pack" JSON. The Caption Engine is completely blind to raw pixels, reading only this verified JSON to generate the four required styles (formal, sarcastic, humorous_tech, humorous_non_tech), effectively eliminating visual hallucinations. A time-bounded Judge loop iteratively refines these captions, locking in the best version before the deadline. To survive strict grading harnesses, the system features atomic writes, ensuring valid output survives a sudden execution timeout at any second. For the purpose of this hackathon, we routed heavy inference to the AMD Developer Cloud via ROCm acceleration, and hosted our Gemma 4 E4B Model on it, and connected to our system. And for the regular video captioning pipeline, the utility tasks were run through Open source models provided by the Fireworks AI. The Gemma Everywhere Upgrade: To unlock next-level capabilities, we unified the stack under Gemma, focusing on enhancing Content Creators. A single Gemma deployment handles vision, captioning, judging, and translation. Because tokenizers are unified, multi-lingual translation happens natively during generation, perfectly preserving sarcasm and humour without context drift. Furthermore, Gemma acts as an Agentic Creative Director. Using multi-turn tool calling, Gemma evaluates the generated text, scans the video frame index, selects the perfect key moment, and autonomously composites custom thumbnails with typography, all safely isolated in a walled-garden architecture so it never bottlenecks the core video captioning pipeline.