AI Smart Crop: Reframe Any Video from Landscape to Portrait

Real-time face tracking

DeepSkim uses a neural network (S3FD) to detect every face in every frame, then TalkNet audio-visual analysis to determine who is speaking. The 9:16 crop window automatically follows the active speaker with smooth, Gaussian-filtered motion — no jumpiness, no manual keyframing.

Split-screen for multi-speaker content

When two people are talking — like in an interview or debate — DeepSkim automatically switches to a split-screen layout showing both speakers. It detects when speaker turns happen and transitions smoothly between single-speaker and split views.

Three crop modes

• Autocrop — AI face-tracking crop that fills the 9:16 frame with the most relevant content
• Blurred background — Full frame visible with a blurred version behind, for a cinematic look
• Black background — Full frame on a clean black background, ideal for text-heavy content

Fallback detection for B-roll

When there are no faces on screen — like during screenshares, slides, or B-roll — DeepSkim uses a 4-tier fallback pipeline (scene detection, object detection, saliency mapping, optical flow) to keep the crop centered on the most interesting part of the frame.