Asking questions about a video, frame by frame

Asking a model about a video by sending it one frame throws away almost everything that makes video video. Grafy sends an ordered set of frames plus a timeline.

A time-ordered sequence

Grafy samples frames across the clip and labels them with timestamps, so the model can reason about what changes and when.

Segments for long clips

Longer videos are grouped into labelled segments, so you can ask things like "compare the mood in segment 1 versus segment 3."

Sound counts too

When there is speech, a transcript joins the context. The result: questions about a clip get answers that actually account for motion, time and audio. Try it on a clip.