Shot 7 of 12 broke and the scene won't cut. Now what?

Every AI film hits this sooner or later. Eleven shots came back fine. One didn't. And because a scene is built as a chain, the broken shot takes its neighbours down with it: the line that runs across it, the look that answers it, the cut that was timed to land on it.

The reflex is to hit regenerate until it comes out right. Sometimes that works. Often it doesn't, because the thing that went wrong wasn't bad luck. It was the model doing exactly what the prompt invited it to do, and it will do it again.

This article walks through one of those shots from The Restoration, a short we made in Huski Studio, and how we got the scene to cut without ever getting a good take of the broken shot. The fix is an old editing trick, a cutaway with an L-cut, and it works in any editor. We'll show it in ours.

The shot that wouldn't behave

The scene is two shots long. A robot painter, Unit 7, stands beside its paintings at a small gallery show. In the first shot, S02, a visitor walks in, glances at a landscape, looks at the robot, shakes his head and leaves. Over it, a narrator reads a review of the robot's work:

"One review said its work was everything art has already been — and nothing it hasn't."

The narrator is off screen. Nobody in the frame says anything. The model lip-synced the visitor to the narrator's line anyway.

We regenerated it three times on the same model. For the last take we hardened the prompt: it said "mouth closed" or "lips sealed" five times, three of them about the visitor, and ended with "Neither character speaks on screen at any point". The narrator's line was labelled "voiceover only, no on-screen speaker". He still mouthed it every time. Our automated clip review, which watches each clip for problems, passed the take as fine to ship. That's worth knowing too: a check that is looking for artefacts won't necessarily notice a performance that is wrong.

Every one of those takes was billed as a full generation, and a fourth would have been billed the same way for the same result.

Why regenerating doesn't fix it

A video model that generates sound and picture together treats a line of dialogue in the prompt as something to perform. If there is a face on screen and a line in the prompt, the most likely video it can make is that face saying that line. Words like "off screen" and "not lip-synced" lower the odds; they don't remove them.

So, regenerating is re-rolling a loaded die. The only reliable ways to stop the model lip-syncing a line are to keep the line out of the prompt, or to keep faces out of the frame. If you're prompting a model by hand, the first one is worth doing from the start: generate the picture without the narration and lay the voice on afterwards.

But suppose, like us, you already have the take, the voice in it is good, and only the picture is wrong. Then this isn't a generation problem any more; it's an editing problem, and editors have been fixing exactly this for a hundred years.

Separate the picture from the sound

The narration in S02 is fine. The first two seconds of picture are fine too, since the visitor walks in and the narrator hasn't started yet. What's broken is the picture from about 2.3 seconds onward, while the line plays.

An editor's answer is to keep the sound and replace the picture underneath it. That's what split edits are for:

• An L-cut lets a shot's sound run on after the picture has cut to the next shot. On a timeline, the audio sticks out past the picture like the foot of an L.
• A J-cut is the reverse. The next shot's sound starts before its picture appears.

Both are usually taught as ways to smooth a conversation, but they are just as useful as repair tools. An L-cut on S02 could end the visitor's picture before he starts "talking", while the narrator carries on.

First attempt: the split edits went the wrong way

Our first try put an L-cut on the shot before S02 and a J-cut on the shot after it, hoping to squeeze the visitor out from both ends.

It did the opposite. The L-cut on S01 ran S01's sound over S02 and shortened S01's picture. The J-cut into S03 started S03's sound early and shortened S03's picture. S02 kept every frame. Following the rule in the note above: to take picture away from S02, S02's own sound has to be the part that runs long.

Second attempt: an L-cut needs somewhere to land

So the split belongs on S02 itself, as an L-cut. How long? The narrator starts 2.35 seconds into the clip and the clip is 7 seconds, so hiding everything after that point means an L-cut of almost 5 seconds.

That's where the next shot gets in the way. With a 5-second L-cut straight into S03, the narrator's S02 line would still be playing when S03's own narration begins:

S03's narrator comes in 1.44 seconds after its cut. The longest L-cut that doesn't talk over it is about 2 seconds, which would still show the visitor mouthing the line for nearly three.

The traditional fix is a cutaway: a short shot of something else in the scene that the sound can play over. It's what documentary editors use to hide a jump in an interview. Here it has a second job: a shot with no faces in it can't lip-sync anything.

The obvious cutaway was the thing the visitor is looking at, the robot's paintings on the wall.

Making the cutaway

In Huski Studio, you can add a shot straight from the Edit timeline, between the two clips where it will go. That opens the Shots Breakdown with a new shot waiting to be described.

The new shot, S02-1, inherits the scene's location, so it is set in the same gallery with the same light. We made sure no characters were in it.

Then the clip itself. The prompt describes an empty wall and ends with "No music." There is no dialogue, so the model has nothing to perform, and no score that would fight the narration playing over it.

The cutaway was usable on the first take. Because it is a second shorter than S02, it cost less to generate than any one of the three rerolls had.

The L-cut

With the cutaway in place, the fix is one setting. Click the marker between S02 and S02-1 and set the audio overlap to an L-cut of 5 seconds.

S02's picture drops to its first 2 seconds, which is the visitor walking in and glancing at the wall. The narration keeps playing over the paintings. S03 starts after the cutaway ends, so its own narration never overlaps. We also softened the cut into the cutaway with a half-second dissolve, and the visitor is gone before the narrator's first word.

It also reads better than the original shot would have. The visitor looks at the paintings and the camera looks with him, while the review describes what we are looking at. The cutaway wasn't in the script, but it's the kind of shot an editor would ask for anyway.

When the cutaway is the wrong fix

This worked because the broken part of the shot wasn't carrying the story. The narration carried it. A cutaway can't save a shot whose performance is the point, like a reaction the scene hinges on or a line a character has to be seen saying. Some other limits:

• The scene changes. It now has three shots and a different rhythm. That's usually fine for a narrated scene, less so in the middle of a conversation.
• The split costs picture. An L-cut of 5 seconds takes 5 seconds of picture out of the shot, so the shot has to have something usable before the problem starts. S02 had 2 seconds, which is about the minimum that still reads.
• The cutaway has to exist. It is one more generation, and it has to match the scene's location and light. That's where a project's saved locations do the work, and it's still worth checking the frame before you cut to it.

If a cutaway doesn't fit, here are the other fixes we reach for, cheapest first:

1. Play it over black. The black-screen transition above, with an L-cut onto it. Free, but it stops the picture.
2. Trim it. If the problem is at the start or end of the clip, drag the clip's edge in the Edit timeline.
3. Cut out the middle. A gap in the middle of a clip can be removed too, but that removes its sound along with the picture.
4. Fix the frame, not the take. If the keyframe is what's wrong, edit it and regenerate from the corrected frame.
5. Edit the clip. For a short section of a clip, the clip editor can re-render just that range from an instruction and keep the original audio.
6. Try a different model. If one model keeps making the same mistake, it's a better bet to try another than to reroll the first.

Try it

Cutaways and split edits are ordinary editing, and every serious editor has them. What's different in Huski Studio is that the scene, its locations and its edit live in one project, so the cutaway comes out in the same gallery and light as the shots around it, and the L-cut compiles into the finished film without exporting anything.

If you have a project open, the Edit timeline is on the Post-Processing step. The clip audio tracks show every split as it will play, and compiling the section renders it exactly. You can start a new project from the Projects page.