Loading tool…
A centre crop of a widescreen video often cuts off the person who is talking. This tool finds faces and people in your clip, plans a smooth camera move that keeps them in frame, and lets you correct any moment by dragging the crop. The result is a vertical, square or 4:5 MP4.
Processing locally on your device
Overview
Cropping a 1920 × 1080 frame to 9:16 leaves a window only 608 pixels wide, under a third of the picture. Where that window sits from moment to moment decides whether the clip works. Smart Reframe samples the video five times per second and looks for the subject with MediaPipe’s BlazeFace detector (about 0.23 MB). In Auto mode, if no face is found in a frame, it falls back to EfficientDet-Lite0 person detection (about 7.3 MB, loaded only when needed). When several people are in shot, it prefers the largest, most confident one that is closest to where the camera was already looking. Both models are Apache-2.0 licensed and run on your graphics card through WebGL, or on the CPU with WebAssembly.
Raw detections jitter, so they are turned into a camera path in a few steps. Gaps are filled between detections, and a dead zone lets the subject move a little before the camera reacts. A zero-phase Gaussian filter (the Smoothness slider) removes shake without adding lag. A pan speed limit, measured in crop widths per second, keeps the camera from whipping across the frame. The path is then reduced to a handful of keyframes joined by smooth ease-in, ease-out curves. These are drawn under the timeline, where you can click any keyframe to jump to it.
No automatic system gets every shot right. If two people sit at opposite edges, or the action matters more than the face, pause on that moment and drag the crop window (or use the arrow keys). That pins a manual keyframe, and nearby automatic ones give way to it. You can delete manual keyframes one at a time, clear them all, or switch to People or Centre mode and re-analyse.
Export uses FFmpeg compiled to WebAssembly. The keyframes become one crop expression that FFmpeg evaluates on every frame, so the output follows exactly the path shown in the preview. The result is re-encoded to H.264 with AAC audio. It keeps the native crop size (for example 608 × 1080 from a 1080p source) or, if you choose, scales to 1080 × 1920, 1080 × 1080 or 1080 × 1350. Analysis of a one-minute clip usually takes tens of seconds. Encoding typically runs at about real time or slower in a browser.
| Source | 9:16 crop | 1:1 crop | 4:5 crop |
|---|---|---|---|
| 1920 × 1080 | 608 × 1080 | 1080 × 1080 | 864 × 1080 |
| 1280 × 720 | 406 × 720 | 720 × 720 | 576 × 720 |
| 3840 × 2160 | 1216 × 2160 | 2160 × 2160 | 1728 × 2160 |
| 1440 × 1080 (4:3) | 608 × 1080 | 1080 × 1080 | 864 × 1080 |
Drop in a horizontal video (MP4, MOV or WebM).
Pick 9:16, 1:1 or 4:5. Analysis starts automatically and finds faces or people.
Play the preview. Adjust Smoothness and Max pan speed until the camera moves the way you like.
Pause anywhere the framing is wrong and drag the crop window to pin a manual keyframe.
Optionally trim the range and turn on HD scaling, then export the MP4.
Use cases
Turn a widescreen interview into vertical clips for Reels, Shorts or TikTok. Where the camera should switch to the other speaker, pin a keyframe on them.
Keep a presenter in frame as they walk across a stage, so highlights work in mobile feeds on Instagram and LinkedIn.
Make 1:1 or 4:5 versions of a YouTube video for feeds where vertical-ish formats take up more screen space than 16:9.
Convert widescreen archive videos into Shorts without keyframing every shot by hand. Let the tracker do the first pass and fix only the misses.
Pro tips
Answers to common questions about the smart reframe.
The camera holds its last position, or eases between the positions before and after the gap. If nothing is found in the whole clip, it stays centred. Either way, you can drag the crop to set keyframes by hand.
It follows one subject at a time. It prefers the biggest face and sticks with it rather than jumping around. To switch to the other speaker, pause where they start talking and drag the crop to them. That creates a manual keyframe the camera eases to.
Yes. The same keyframes and the same easing curve drive both the preview and FFmpeg’s crop expression. Positions can differ by a pixel or two, because H.264 needs even crop offsets.
Cropping changes the picture itself, so the video must be re-encoded. The audio is re-encoded to AAC so it works in every MP4 player. At the native crop size, quality is visually unchanged (H.264, CRF 20).
Analysis checks five frames per second, so a one-minute clip means about 300 detections, usually done in tens of seconds. Encoding typically runs at about real time or slower, depending on the resolution and your computer.
No. ToolEnhancer processes files on your own device inside the browser. They are never sent to, stored on or seen by our servers.
Yes. The smart reframe is free to use with no sign-up, no watermarks and no daily limits.