peek
frame · selector

Find the frames
that matter most.

PEEK ranks every frame in a video by how useful it would be for a captioner — no caption, no prompt, no labels at inference. Pick how many you want; we hand them to a vision-language model and let it talk.

01

Upload & select

12345678
Run on
Compare end-to-end speed. GPU uses ZeroGPU on the Space.
02

Review the picks

03

Caption

Prompt
SmolVLM2-2.2B-Instruct will write one caption per frame set.
PEEK

Uniform