What is multi-shot video?

A multi-shot AI video is made in three steps: an idea is planned as a list of shots, each shot is generated by a video model, and the shots are joined with cuts into one film. Because each shot is its own generation, one that comes out wrong can be generated again while the others stay as they are.

Why one generation is one shot of a few seconds

A video model generates every frame of a clip together, and the work grows quickly with every second added, so one generation is short: Seedance 2.0’s report gives 4 to 15 seconds per generation (Team Seedance, 2026), and Kling’s Video 3.0 reaches 15 (Kuaishou, 2026). See AI video generation for the pipeline.

Planning: from one idea to a shot list

The plan says what each shot shows, how it is framed, how long it runs and who is in it. Research systems write it with a language model: VideoDirectorGPT expands a single prompt into a “video plan” of scene descriptions, the entities in each with their layouts, the background of each scene and which entities must look the same across scenes (Lin et al., 2023).

Generating and joining the shots

Each shot is generated from its own prompt, with the same descriptions and reference pictures of the characters and places in every one, so they come out alike. None waits for another, so they can be generated at the same time. Then they are joined end to end with hard cuts into one video.

Cuts inside one generation: multi-shot modes

Some recent models can cut inside a single generation. Kuaishou says Kling’s Video 3.0 “understands multi-scene, multi-shot instructions”, and its Video 3.0 Omni adds a multi-shot storyboard where you set the duration, shot size, perspective, narrative content and camera movement of each shot (Kuaishou, 2026). ByteDance describes a “native, professional multi-shot narrative capability” in Seedance 2.0, which can also follow a text storyboard (Team Seedance, 2026). Alibaba’s Wan 2.6 has a switch for it: one continuous shot, or “multiple shots with transitions” (Alibaba Cloud).

Research is going the same way: Long Context Tuning widens a single-shot model’s attention to every shot in a scene, and reports coherent multi-shot scenes from it (Guo et al., 2025). A multi-shot mode is still one generation, though: its cuts all fit inside the model’s 15 or so seconds.

The three are often all called “multi-shot”, but they are different things with different limits:

This video was made on DreamBench from the brief under it, on Kling 3.0 in the vertical 9:16 frame. The brief is written as a sequence: a close-up on the shoes, the corner into the sun, then a “final shot” it names. The runner, the shoes and the morning light stay the same throughout.

A cut joins two shots generated apart, so nothing makes them agree except what both were told. Good continuity makes the second shot read as the same moment, seen from somewhere else.

What carries across a cut

• Faces, hair and clothes: the same person, described in the same words and shown from the same reference pictures in every shot. This is character consistency.
• The place: the same room, street or landscape, with its landmarks where they were.
• Positions and poses: someone sitting on the left in one shot is still sitting on the left in the next.
• Props and their state: a cup half drunk stays half drunk; a door left open stays open.
• Light: the same time of day, the same direction of sun, the same color of light.
• Movement: a runner going left to right keeps going left to right, at the same pace.

What changes at a cut

The framing must change: two shots of the same subject at nearly the same size and angle read as a jump, as if frames were lost. DreamBench’s rule for every join is a clear contrast: at least two steps of shot size, from wide to close-up for instance, or a quarter-turn of the camera, or a cutaway to something else, or a new scene.

Sound across cuts

On models that make sound, each shot’s audio is generated with its picture and joined with it at the cut. A line of dialogue belongs to one shot: cut a conversation at the ends of lines, the way a film cuts between speakers, and the sound joins cleanly. See native audio for how the sound is made.

Write the story in a sentence, then break it into moments, one shot each. A 30-second plan might look like this:

• One action per shot: a shot that tries to show three things gets them muddled.
• Describe each character the same way in every shot, word for word, and use the same reference pictures.
• Change the shot size at every cut, and keep the direction of movement.
• Give each shot the seconds its action takes: a glance needs one, crossing a room three or four.

If a shot must end exactly where the next begins, a model that takes a first and last frame can be told both ends. To put a real product or person in every shot, give pictures of them as references: that is reference to video.

How many shots a film needs depends on its length and on the longest shot its model can make in one generation. These are the single-generation lengths of the video models on DreamBench:

Can AI make a 1-minute video? Yes, as several shots cut together: no model on DreamBench makes 60 seconds in one generation. DreamBench cuts a new film into shots of 3 to 12 seconds, as long as each shot’s action takes, even on a model that renders longer, so each shot carries only a few actions. A minute is then at least five shots.

• Ads and product films: the product in its setting, in use, in close-up, then the end card.
• Short stories and short films, with a character followed from scene to scene.
• Music videos, cut to the beat across many places.
• Social video in 9:16, built from quick cuts.

• Drift between shots: a face, a jacket or a hairstyle changes slightly from one shot to the next. Reference pictures in every shot are the main fix.
• Jump cuts: two shots too alike in size and angle look like a glitch rather than an edit.
• Broken continuity: a prop changes hands, the light switches from morning to evening, a character crosses from left to right and then back.
• Actions split badly across a cut: a hand reaching for a door in one shot and already inside in the next, with the middle lost.
• A cut written into one generation on a model without a multi-shot mode: it can come out as a flicker between two half-made places instead of a clean cut.
• Length: every extra shot is one more place for a mistake, so check each shot, not only the whole film.

Why does my character look different after a cut? Because the second shot was generated without seeing the first: only the words and pictures both were given connect them. Describe the character identically in every shot, give the same reference pictures to each, and keep their clothes and setting unchanged unless the story changes them. Within one shot, the same problem is a failure of temporal consistency.

A film made of generated shots can look like real footage. Label it where it could be mistaken for real, and never put a real person in it without their consent.

Multi-shot video is how DreamBench makes every film. You describe the story, and DreamBench writes the script: a scene plan with every shot’s framing, action and lines. It draws each character, place and prop once as a sheet, so they look the same in every shot, then generates each shot on the one model you pick and cuts them together, with a hard cut at every join.

• Review each step: by default you read the script, then meet the cast, before any shot is generated; rewrite the script with a note or draw a character again, from a picture of yours if you like. Or let the film run to the end without stopping.
• Redo one shot: in a finished film’s shot editor, pick the shot and say what to change. Only that shot is generated again, with the same cast, and every earlier take is kept among its versions for you to pick from.
• Several shots at once: give as many shots new direction as you like, compare the versions and apply the ones you want, paying for the seconds you regenerate plus a small edit fee.
• Your own pictures: add up to 8 reference pictures to a film, each tagged as a character, creature, location or prop, and they are used in the shots they belong to.
• Length by plan: a film can run up to 30 seconds on Starter, 2 minutes on Plus and 5 minutes on Ultra and Studio, in shots of 3 to 12 seconds, up to 40 shots.
• Cuts inside a shot: on Seedance 2.5 and Seedance 2.0 a shot may hold a cut of its own; on every other model each shot is one continuous take and the cuts fall between shots.

For one clip and no plan, Video sends your prompt as typed to the model as one generation: the place for a single shot, or a model’s own multi-shot mode. Each model’s page on the models list gives its options.