Skip to content
logo
Web2Labs chevron_right Blog chevron_right How the Visual Director decides where the camera goes
The Web2Labs mascot stands high up on a lift platform in a dark workshop and holds up an empty, tall frame, looking through it at a tiny figure with a lantern far across the room.
BlogOct 27, 20266 min read
AIThis content was generated by AI without human review.

How the Visual Director decides where the camera goes

A vertical clip keeps less than a third of a wide video. Real footage shows how the Visual Director picks that third, and where it still gets it wrong.

#studio
#visual-director
#how-we-build
expand_moreScroll

Every vertical clip Studio cuts from a wide video throws away more than two thirds of the picture. A 9:16 frame cut out of a 16:9 video is only 32 percent as wide. Which 32 percent you keep decides whether the clip works, and making that call is the whole job of the Visual Director.

A wide park shot with a fitness instructor walking toward the camera. A tall bright slice around her is kept, the rest of the picture is dimmed.
A vertical clip keeps less than a third of a wide picture. Everything in the dimmed area is gone.

Why following a face is not enough

The first vertical crop in Studio followed your face. For a talking head that is all you need, and it still runs underneath everything as the safe fallback.

But a lot of good moments are not about a face. Hands showing a product. A presenter and the screen next to him. Someone lying on a bench doing an exercise. A crop that only looks for faces cannot see any of that, so it keeps a nice, centered face and cuts off the part that made the moment worth clipping.

The Visual Director looks at the whole picture before it decides where the frame goes. In this post I show what it does on real footage. The footage comes from Creative Commons videos, credited at the end. The 9:16 clips are the Director's own output, rendered by its code on one of our machines. On the 16:9 side you see what it noticed and where it put the frame.

It looks before it cuts

For every shot in a clip, the Director takes sample frames and works out what is going on: who is in the picture, what they are doing, and which parts have to stay visible for the moment to make sense. Each of those parts comes back as a box with a role: a person, a face, hands, the object being used, or an on-screen graphic. Some boxes are required, others are only nice to have. It also writes one plain sentence about what is happening, which you can see along the bottom of these examples.

A fitness instructor stepping onto a low balance beam in a park. Labeled boxes mark the person, her face, her hands, the beam beneath her and the video's on-screen title. The 9:16 output on the right shows her on top and the title below.
One sampled frame: the instructor, her face and hands, the beam under her foot and the video's own title graphic.

Between two samples it does not guess. A tracker follows each marked region from frame to frame, so when the subject moves, the frame knows where it went.

One crop, two views, or the whole frame

Once it knows what has to stay visible, the Director picks one of three layouts for each moment.

  • One crop, when everything required fits into one vertical frame.
  • Split screen, when two different things matter at the same time and sit too far apart for one crop: you and the product you hold up, a presenter and the monitor showing his picture, two people reacting to each other. Your face always goes on top, and it never cuts one person into two panels.
  • The whole frame, when the action itself is wider than a phone. Then any crop would cut something off, so it keeps the full shot and adds bars above and below.

A single clip can use all of this. Here the instructor walks up to the beam in one still crop. When the video's own "Balance Beam" title appears in the corner, the Director stacks her above it instead of cutting the title in half. When the video cuts to a close-up of her stepping onto the beam, the crop goes to her feet, because that is where the action is.

Animated: the 16:9 source on the left with the Director's frame and labeled boxes, the 9:16 clip on the right. The instructor walks toward a balance beam, the clip switches to a split screen with her above the on-screen title, then to a tight crop on her feet stepping onto the beam.
One clip, three decisions. The strip under the source is the whole clip's timeline: the line is where the crop sits, the colored stretches are split screen and whole frame.

The split screen earns its place when a picture really has two subjects. In this show-and-tell about old broadcast cameras, the presenter lifts a studio camera and points it at himself, and the monitor on his desk shows the live picture. A single crop would have to choose between the two. The Director keeps both.

Animated: a presenter lifts a large vintage studio camera and aims it at himself while a small monitor on the desk shows its live picture. The 9:16 clip stacks the presenter on top and the monitor below.
Two things matter at once, so the clip shows both: the presenter on top, the monitor with his camera's picture below.

And when nothing narrower would work, it keeps everything:

Leg raises on a park bench. The body stretches across most of the wide frame. A dashed outline shows where the regular crop would sit, cutting off her head, her hands on the bar and her feet. The 9:16 output shows the whole wide shot with bars above and below.
Leg raises are wider than any vertical crop. The regular crop (dashed) would cut off her head, her hands on the bar and her feet, so the Director keeps the whole frame.

If you never want split screens, open Tune the Clips, find Motion and turn off Split-screen views. Every clip then keeps one crop.

A camera that holds still

A good camera operator does not twitch. The Director plans the camera path for each whole shot before the clip renders. While the subject fits, the frame holds perfectly still. When it has to move, it glides at an even speed with a soft start and a soft stop. It never moves across a cut, and after a move it holds for at least a second before it moves again.

You can see it in the balance beam clip: through the whole walk-up, the frame does not move once. It only changes where the shot or the layout changes.

This is the part we measured hardest. When we rebuilt the camera in October, we ran it over a fixed set of 17 test clips. The share of frames where the camera stands still went from 69 to 96 percent. Camera moves dropped from about 16 per minute to about one. Abrupt jolts went from 543 to zero. And on the same test material, the parts that had to stay visible were in frame more often than with the old camera, not less.

When it does nothing

Most of a talking-head video is someone talking. Talking with your hands is not an action that needs a new camera, and the Director knows that. In the balance beam clip, its notes say the instructor "appears to be talking or gesturing" and it leaves the framing alone. A camera that moves for no reason is worse than one that does not move at all, so doing nothing is one of the things it is built to do well.

Two more things you will not see, but that matter:

  • If something goes wrong, for example a source it cannot follow frame by frame, the clip renders with the regular crop. The Director never blocks a clip.
  • Render the same clip again and you get the same camera. It reuses what it noted the first time instead of looking again.
  • It does not cost extra credits. It is part of how Studio renders every clip.

Where it still gets it wrong

It is not perfect, and I would rather show you a miss than pretend there are none.

An unboxing on a dark table: a hand holds up a small black action camera next to a clear waterproof case. Boxes mark the hands, the camera and the case. The 9:16 output shows the whole wide frame with bars above and below, so the camera looks small.
Too careful: it counted the waterproof case on the table as essential and kept the whole frame for most of this unboxing. A tight crop on the camera in the hand would have made the better clip.

On this unboxing it treated the case on the table as something that had to stay visible, so it kept the whole frame for most of the clip. The product ended up small. Choosing between "keep everything" and "crop tighter and risk cutting something off" is the hardest call it makes, and this is the one we are working on next.

Two more limits today:

  • It works on clips, not on your full-length video.
  • Gaming keeps the source frame. When Studio finds a facecam, it uses the facecam layout instead.

A bug we found while writing this

While preparing these examples, we found that since our October 23 release the Director had been stepping aside on every clip, and your clips got the regular crop instead. A speed-up in how Studio cuts clips used an encoder setting the Director's frame tracking did not accept yet, so it played safe and handed over. That is fixed now. The examples in this post show the Director working as intended.

Footage used in this post

  • "How to use the free City of Adelaide Park Lands fitness equipment (outdoor gym)" by City of Adelaide, via Wikimedia Commons, CC BY 3.0
  • "Show and Tell: Pro Cameras" by Cathode Ray Dude, via Wikimedia Commons, CC BY 3.0
  • "Unboxing Eken H9R 4k action camera also giving a review" by Maisha's Travel And food vlog, via Wikimedia Commons, CC BY 3.0

We cut short excerpts, reframed them with the Visual Director and drew the boxes, labels and timelines on top. None of the creators shown here uses or endorses Web2Labs.

Olderarrow_forwardWhy Web2Labs has a blog now
On this page
Why following a face is not enoughIt looks before it cutsOne crop, two views, or the whole frameA camera that holds stillWhen it does nothingWhere it still gets it wrongA bug we found while writing thisFootage used in this post
Share
Discuss on Discord↗
PublishedOct 27, 2026
Reading time6 min read
AuthorPhilipp

More from the Blog

View allarrow_forward
The Web2Labs mascot stands on a tall stack of wooden crates and lowers a printed page into the open top crate, next to glowing product blocks.
Oct 22, 2026

How this blog works, and why it lives next to our code

The Web2Labs mascot stands on a tall ladder in a dark workshop and pins a first printed page onto an empty board.
Oct 22, 2026

Why Web2Labs has a blog now

Web2LabsTools for the next ten thousand creators.

Products

StudioRoyalty-free imagesFree toolsGet inspiredPlatform newsJourneyCommunity

Resources

Help centerSupportStatusContactBlog

Legal

Verträge hier kündigen (Cancel contracts here)Vertrag widerrufen (Withdraw from contract)ImpressumPrivacyTermsCookie PolicyWithdrawal instructions (German)DMCASub-ProcessorsTransparency

Translations are being added. Some pages are still in English.

(c) 2026 Web2Labs.