VoxFrame: Multimodal Narrative Synthesis Engine

Vercel
application badge
Created by team CognitiveFrames on July 10, 2026
Video Captioning

VoxFrame is a multimodal narrative synthesis system built for the AMD Developer Hackathon ACT II Track 2 challenge. It addresses common video-captioning failures such as hallucinated details and inconsistent tone by combining structured scene grounding with style-aware caption generation. The pipeline downloads the source video, extracts adaptive keyframes with FFmpeg, optionally transcribes speech with Groq Whisper, and then uses AIMLAPI-connected Gemini models to build a verified scene summary from the images. That scene summary is used to generate four distinct caption styles: formal, sarcastic, humorous_tech, and humorous_non_tech. A local audit and rubric-based scoring pass check for realism, tone compliance, and style fit, while low-scoring outputs can be repaired or refined before final export. The project is packaged as a Docker image and served through a FastAPI-powered web dashboard for easy local and cloud deployment.

Category tags: