Cutting without jumps
Lesson 14 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
The 180° rule
When there are two people in a scene, an imaginary line runs between them: the axis. As long as every camera stays on one side of it, Mira is always on the left and Lena on the right, and the viewer doesn’t get lost. If one camera jumps the axis, in the next shot the characters seem to have swapped places. Every editor knows this 180° rule, but AI doesn’t: the model generates each shot separately and easily puts the camera on the other side. So state in the prompt who is where: “Mira on the left of frame, Lena on the right, camera on the street side”. And keep it identical for every shot of the dialogue.
Screen direction
If Mira runs left to right in the first shot, she should run the same way in the second. Then the brain reads it as one journey: she keeps running. Flip her to right-to-left in the second shot, and the viewer decides she changed her mind and is running back. Films shoot chases this way: the “good guys” always move one way, the “bad guys” the other. AI doesn’t remember this, so write the direction explicitly too: “runs from left to right of frame”. If the story needs a turn, show it in the frame instead of hiding it behind a cut.
Chaining frames
The most reliable way to join two shots without a jump is a chain. When shot A is ready, take its last frame and make it the first frame of shot B in image-to-video mode. Shot B then starts from exactly the pose, light and place where A ended. In prompt B describe only the new action; the model takes the look and place from the image. That’s how long scenes are built from short generations. One downside: errors also pass down the chain, so check each shot before taking a frame from it. Many models can also do the reverse: set both the first and the last frame, and the model fills in the motion between them.
The rule of six
Walter Murch is the editor of “Apocalypse Now” and a three-time Oscar winner. In his book “In the Blink of an Eye” he proposed six criteria for a good cut and even gave them weights. Emotion: 51%, the cut should carry what the viewer feels at that moment. Story: 23%, does the cut move the plot forward. Rhythm: 10%. Eye-trace: 7%, where the viewer was looking before the cut is where the important detail should land after it. And only 9% for the rules of the screen plane and space, like the 180° axis. The takeaway for AI videos: a small jump is forgiven if the cut hits the emotion, while a technically perfect but empty cut won’t save anything.
The cutaway
Sometimes two shots of Mira almost match, but not quite: the hair lies a bit differently or the light has shifted. Redoing is expensive, and cutting straight between them shows the jump. An old editing trick is the cutaway: a one-to-two-second detail goes between them. A close-up of the hoodie fabric, sneakers on asphalt, the city from above. While the viewer looks at the detail, the brain “resets” the image, and the small difference gets lost. For a clothing brand it’s a double win: the fabric detail also sells the material. Generate a couple of these details for every video in advance; they almost never glitch.
No morphing
Many AI tools offer “smart” transitions where one scene flows into another and a face smoothly turns into another face. In 2023–2024 it looked like magic; now viewers read morphing as a caption saying “look, a neural network made this”. For an ad that wants to look filmed, the best transition is a straight cut on action: a door opens, a hand reaches, a person jumps, and the cut lands mid-motion. Leave morphing for music videos where it is the style, and use it on purpose, not because the button was right there.
The storyboard
Before generating, pros draw a storyboard: one panel per shot, with shot size, action and timing. It’s cheap, a sheet of paper or a note, and it saves loads of credits because you generate only the shots you need. Look at the storyboard for Mira’s 15-second video. The wide shot sets the place, the medium shows the hoodie on the body, the detail sells the fabric, the close-up gives emotion, the jump is cut on action, and the end card is the logo and the drop date. Each shot has its own job, and together they tell a story instead of just showing pretty pictures.
Check yourself
- In shot 1 Mira runs left to right, in shot 2 right to left. Viewers write: “wait, did she go back?” How to fix it?
Correct answer: Regenerate shot 2 with “runs from left to right of frame”. The brain reads one screen direction as one journey. It must be set explicitly in the prompt. - In Mira and Lena’s dialogue, Mira is on the left in one shot and suddenly on the right in the next. What rule is broken?
Correct answer: The 180° rule: the camera jumped the axis. Cameras must stay on one side of the line between the characters. Then their places in frame don’t change. - Mira wants the jump from shot A to continue into a landing in shot B with no jump in the image. How?
Correct answer: Use A’s last frame as B’s first frame and describe only the landing in the prompt. Frame chaining carries the pose, light and place exactly from the end of shot A, and the motion continues without a jump. - Two shots of Mira almost match, but her hair lies slightly differently. Redoing takes long. What goes between them?
Correct answer: A 1–2 second cutaway: a close-up of the hoodie fabric or sneakers. A detail “resets” the image in the viewer’s head, and the small difference gets lost. The fabric close-up also sells the hoodie.
← One face for the whole video Edits instead of regenerating →