--- title: "Auto reframe explained: how face tracking crops 16:9 to 9:16" description: "What happens when software crops a widescreen recording to vertical, how face tracking picks the subject, and when it gets it wrong." summary: "Auto reframe crops a widescreen frame to vertical by deciding, frame by frame, what to keep. A face detector locates people, a tracker follows them across time, and the crop window moves toward whoever is speaking, smoothed so it pans rather than jumps. Roughly two thirds of the original width is discarded." url: "https://www.gpt-video.com/blog/auto-reframe-explained" published: "2026-08-25" updated: "2026-08-25" author: "Alessio Battagliero" topic: "AI video editing, explained" keywords: "auto reframe, face detection, aspect ratio, video editing" --- # Auto reframe explained: how face tracking crops 16:9 to 9:16 Auto reframe crops a widescreen frame to vertical by deciding, frame by frame, what to keep. A face detector locates people, a tracker follows them across time, and the crop window moves toward whoever is speaking, smoothed so it pans rather than jumps. Roughly two thirds of the original width is discarded. ## The problem: two thirds of the frame has to go A widescreen recording is 16:9. A vertical clip is 9:16. Converting one to the other is not a resize — it is a decision about what to discard. The arithmetic is stark. Take a 1920x1080 frame and crop it to 9:16 at full height: the resulting window is 608 pixels wide. About two thirds of the original width is thrown away, permanently, on every single frame. So the only interesting question is *which* third to keep, and the answer changes constantly as people move, lean, gesture and take turns speaking. Cropping to a fixed centre works only for a single motionless speaker perfectly centred in shot, which describes almost no real recording. Everything else needs the crop window to move. Doing that by hand means keyframing a position track across the whole clip. It is not difficult work. It is just slow, and slow enough that most people skip it and accept a fixed crop that cuts off half the conversation. ## How the automatic version works Three stages, running in sequence over the footage. **Detection** finds faces in individual frames. A detector scans each frame and returns bounding boxes for the faces it is confident about — position and size, nothing more. **Tracking** connects those detections across time. A face in frame 100 and a face in frame 101 need to be understood as the same person, so the system can follow one subject rather than jumping between whoever scored highest this frame. Tracking is also what carries the subject through the frames where detection fails — a turn of the head, a moment of shadow — instead of dropping them. **Smoothing** turns the resulting path into camera movement. A crop window that follows the raw tracking data exactly would jitter, because detection boxes wobble frame to frame. The path is smoothed so the crop drifts and settles like an operated camera, and only moves when the subject genuinely moves. :::key Why smoothing is the part that makes it look professional - Raw tracking produces a crop that vibrates — technically correct, unwatchable - Over-smoothing produces a crop that lags behind the speaker and arrives late - The target is a camera that appears to have anticipated the movement ::: ## Speaker switching, and the split-screen alternative Two people on screen is the case that separates good implementations from bad ones. If the crop simply centres on whoever is detected, it will bounce between two faces at conversational speed, which is unwatchable. The better behaviour is to follow the *active speaker* — using the audio to determine who is talking and holding on them until the turn genuinely changes, with a deliberate pause so a two-word interjection does not trigger a cut. The alternative is not to choose. A [split screen](/features/split-screen-video) stacks both speakers vertically, keeping each in their own frame. This works well when the reaction matters as much as the words — comedy, disagreement, anything where the listener's face is content. It works badly when one person is doing all the talking, because half the frame is then a person listening politely. The rule of thumb: follow the speaker for interviews and interrogative conversation, split for reaction and banter. If in doubt, follow the speaker — it is closer to how a viewer would look at the room. ## When it gets it wrong Predictably, and in ways worth knowing before you trust a batch of clips. **No face to follow.** Screen recordings, slides, product shots, b-roll. Face tracking has nothing to lock onto and will either hold centre or drift. For these, a manual crop is both faster and better. **Profile and back-of-head shots.** Detectors are trained mostly on frontal faces. A speaker turned away can be lost, and the crop will wander toward whoever else is visible. **Fast movement out of frame.** Someone standing up quickly outruns the smoothing, and the crop arrives after they do. **Difficult lighting.** Strong backlight, deep shadow and heavy colour grading all reduce detection confidence, which shows up as a crop that hesitates. **Crowds.** More faces than the frame can hold means constant, arbitrary choices between them. Crop manually. The common thread is that failures are visible immediately, in the first pass, on any clip you actually watch. Which is why the [review step is not optional](/blog/one-hour-weekly-clip-workflow) — checking framing on a phone-sized viewport takes seconds per clip and catches all five of these. ## Resolution: the trap nobody mentions This is the part that quietly degrades output, because it produces no error and no obvious artefact. Crop a 1080p widescreen source to vertical and you are left with roughly 608 pixels of width. Vertical platforms display at 1080 wide. The player upscales, and the result is soft — not broken, just consistently less crisp than the clips it sits next to in the feed. Record in 4K and the same crop leaves about 1216 pixels of width, comfortably above the target, and the clip is genuinely sharp. This is the strongest practical argument for recording at higher resolution than you intend to publish: not for the detail, but for the crop budget. If you are stuck with 1080p sources, the mitigation is to crop less aggressively where the framing allows — a slightly wider shot loses less width than a tight one — and to accept that the output will be soft. ## Where it fits Reframing is one of the four jobs that make up [AI video editing](/blog/what-is-ai-video-editing), and it is the one with the clearest verdict: for talking-head footage with faces to follow, automatic reframing is both faster and steadier than a person keyframing under time pressure. For anything without a face in it, it is the wrong tool. Treat it accordingly. Let it handle the conversational clips, which is most of them, and reach for a manual crop the moment the subject is not a person.