Prism - One clip, four voices
Prism follows a ground once, restyle four ways pipeline: Sample - pull frames evenly across the clip (9 / 16 / 25, scaled to the clip's length). Montage - tile those frames into a single grid image, so the vision model reads the whole clip in one small request. Ground - Gemma-4 studies the montage and writes one detailed, factual description of what happens. Restyle - that single description is rewritten into all four styles in one structured call. Grounding once keeps every caption faithful to the same facts while each one nails its own voice. Gemma-4 does both the vision and the styling - the pipeline is Gemma end to end. For reliability there's a three-tier failover - HuggingFace Inference Providers → Fireworks → an AMD-hosted endpoint - with retries on transient errors, so the live output path never dies.
Category tags:
For Innovators & Creators