
Hawk VCA is a layered video-captioning system built for AMD Developer Hackathon Track 2. It separates understanding what happens in a video from deciding how that information should be expressed. The perception layer watches the clip and creates facts with visible subjects, actions, setting, and boundaries for claims the evidence does not support. The style-rendering layer receives those facts and generates the formal, sarcastic, tech-humor, or non-tech-humor caption without changing facts. Next: perception will transform into an ensemble that combines complete-video and frame-level evidence, and the renderer will be fine-tuned using style examples.
13 Jul 2026