Generate a whole AI scene, shot by shot

Why a scene is harder than a clip

Ask a video model for "two people arguing in a kitchen" and you get a few seconds of two people in a kitchen. Ask again for the next moment and you get a different kitchen. A scene needs the opposite: one room that stays the same, the same two people, and a run of shots that cover the argument line by line.

Scene generation is therefore about continuity inside a sequence. The room's layout, who stands where, which way they face, what's on the table, all of it has to survive from the wide to the close-ups and back.

Screen direction is the part people notice without being able to name. If she is on the left looking right in the wide, she should still be on the left looking right in her close-up, or the conversation stops making sense. Props behave the same way: the glass in her hand in one shot can't be on the table in the next unless someone put it there. A scene generator has to plan these things once for the whole scene. Each shot can't be left to decide them alone.

How to make a scene that cuts

Keep each scene to one location and time. A different room is a different scene. Open with a shot that establishes the space, then cover each line with a shot of the speaker or the listener. Add inserts for the objects the story needs. Change shot size between cuts so the edit reads.

Watch the timing. Each line needs enough screen time to be spoken, and a scene of short shots needs lines short enough to fit them.

• One location and time per scene.
• Establish the room, then move in.
• Cover every line; give reactions their shots.
• Fit each line to its shot's length.

How ShowMaker holds the room and the cast

Locations have reference stills, and characters have character sheets, wardrobe and voices. Every shot in the scene cites them as @references, so the same room and the same people are sent to the model on each shot instead of being re-described.

Scene Flow puts the script and the storyboard on one surface: the scene text, the shots, their frames and their previews side by side. Checks on every shot flag a line too long for its shot, or a frame that shows something the text never mentioned. Shoot with AI fixes a frame from a plain note and shows the credit cost first.

Multishot panels for dialogue scenes

Dialogue scenes are where per-shot rendering struggles, because each clip is generated separately. A multishot panel renders several shots of a scene in one take, cut on dialogue, so the characters and room stay consistent within the take. The take can then be trimmed and broken back into individual shots for the edit.

Scenes can also mix the two. Use a multishot panel for the conversation and separate renders on cheaper models for the establishing wide and the cutaways. A free narrated animatic lets you watch the whole scene before deciding what to render.

Most of what a scene costs comes from regenerating shots that came out wrong. So approve the frames in Scene Flow, listen to the voiced lines and watch the animatic before you render. Then the video you pay for is motion added to pictures you already accept. Spend premium credits on the close dialogue and cheaper ones on the wide and the inserts. That saves your budget for the shots viewers watch most closely.