Why the moving crop looks broken
Most long-form podcasts are two people at a table, filmed wide. Turn that into a vertical video by cropping to whoever is speaking and you get a frame that jumps every few seconds, cuts off the reaction that made the line funny, and occasionally lands on a microphone stand while it decides who to follow. Viewers read this as a bug, because it is one.
The fix is not a smarter chase. It is a different layout: two crops stacked vertically, one speaker on top and one below, so both faces stay on screen for the whole clip. The reaction is visible, the frame never jumps, and the clip reads as a conversation rather than a slideshow. This is the default our split-screen clip workflow produces for a two-person episode.
When to stack, and when to follow one person
Stacking is right when both people are on camera for most of the clip and their faces are comparable in size, which is what a face-to-face setup gives you. It is wrong in a few cases that come up often enough to name:
- The episode is really one person plus a silent host. If the second face is barely present, or the second person barely speaks, a single crop that follows the speaker is cleaner. ClipFlap checks how much of the clip each person talks, using the speaker labels in the transcript, before it chooses a split: a listener who only says "right" twice does not get half the screen.
- More than two people at the table. Stacking two of four guests leaves out the others. When one person is clearly speaking, a crop on that person works; when the table is a real group conversation, the whole shot is the honest frame.
- The camera cuts between close-ups. When the source edit already alternates single shots, keep each shot as a single and change framing on the cut, never in the middle of a shot.
- A moment depends on one face. A long personal story is better held on the person telling it, with the listener's reaction shown only when it is the point.
The decision should be made per camera shot rather than once per clip. A clip that opens on a wide two-shot and continues into a close-up should stack for the first part and go single for the second.
The false positives that fool face detection
Two-person layouts are decided by counting faces, and counting faces is harder than it sounds. Three cases catch out most tools, including early versions of ours:
- A portrait on the wall. A framed photo behind the host is a face that is present in every frame and perfectly still. If a tool treats stability as a good sign, the photo becomes the second speaker and half the screen shows a picture.
- A face on a phone or monitor. A video call held up to the camera is a face that moves, so stillness will not catch it. What gives it away is size: a face on a screen is much smaller than a person at the same table.
- The back of a head. Over-the-shoulder shots produce a confident detection on the interviewer's hair. Nobody wants a stacked clip whose top panel is a neck.
If your clips sometimes come out stacked with a photograph or a blur in one panel, this is why. The fix in our pipeline was to judge candidates relative to each other, comparing jitter and size against the faces that are clearly alive in the scene, rather than applying a fixed threshold.
When the automatic choice is wrong
No automatic rule gets every clip right. In ClipFlap's clip editor you can override the framing for one clip and render it again: keep everyone in the full frame, hold a fixed crop on the left or right, stack two fixed zones, or follow the active speaker, where the frame switches to whoever is talking, turn by turn. Following the active speaker is a choice you make for a clip, not the default: for a face-to-face conversation, the stacked layout keeps the reaction on screen.
Where the hook card goes
On a stacked clip, the on-screen title has nowhere obvious to sit: the top of the frame is a face and so is the bottom. Put it at the seam between the two panels for the seconds it is shown, and it covers neither. This sounds like a detail until you see a title across someone's eyes.
Captions across two speakers
Word-timed captions work the same way on a stacked clip, but the speaker change should be visible in the transcript that drives them. When the source already provides speaker labels, use them. When it does not, the layout decision falls back to geometry alone, which is why the size and stillness checks above matter so much.
For the selection side of the same problem, which moments to pick in a conversation, see how to find the best moments in a podcast.
Framing is one decision among six; the complete guide to podcast clips covers the rest.