🤓 Latest Submissions
StyleReel: Gemma video captions, four voices
StyleReel takes a short video and writes four captions for it, each in a different voice: formal, sarcastic, tech-humor, and everyday-humor. The catch is that all four stay true to what is actually on screen. Only the tone changes. Here is how it works. It pulls a handful of smart frames and transcribes the audio itself (Fireworks retired its speech API, so we run Whisper locally, which is why our captions actually get the dialogue). Then one model watches those frames, writes down what is happening, and checks its notes against the frames a second time so nothing gets invented. Four separate writers each take that fact sheet and find their voice, best of five drafts. A different model grades every draft for accuracy and keeps the strongest one. Because all four captions come from the same checked description, they never argue with the video or with each other. The model that watches the video is Gemma 3, self-hosted on an AMD Radeon PRO W7900 with vLLM and ROCm. We did not just call Gemma over an API. We ran its multimodal vision ourselves on AMD hardware, and head to head it matched our best cloud vision model on accuracy and style. We also cared about never failing. A valid results file exists from the first second and refreshes after every clip, so a crash or a timeout never leaves you with nothing. Every style is always there, the JSON is always valid, and it exits clean. We threw dead URLs, odd inputs and malformed data at it and it kept working. On our own test set it scores about 0.93 on style, and a blind judge tells all four voices apart every time.
13 Jul 2026
