Writing a scene with people and dialogue

One render can run up to 30 seconds on Seedance 2.5 and 15 on Seedance 2.0, and the model cuts between camera angles inside a single render. For anything longer than a few seconds, write the scene as a short shooting script. On Seedance 2.5, three to five hundred words is normal for a 20 to 30 second scene: the model draws only what it is told. Seedance 2.0 does better with shorter, plainer prompts: give it the shots and the lines, not a whole script.

The shape that works best, in this order:

• Who is who: one line per picture (Image 1 is …).
• What happens: one or two sentences, the style, and how the camera moves across the scene.
• The shots. On Seedance 2.5, with whole-second times and no gaps: Shot 1 (0–4s): [framing, camera move] what happens. Seedance 2.0 ignores times, so number the shots there instead (Shot 1:, Shot 2:). Two to six seconds of action each.
• Each line of dialogue inside the shot that speaks it, as Name (how it is said): "the line".
• Requirements: the look; each person's age, build and exact clothes, and that faces and clothes stay the same throughout; each speaker's voice; the sound.

Image 1 is Marcus. Image 2 is Lena.

Marcus and Lena argue in a diner booth late at night until Lena slides a key across the table. Grounded cinematic live action; the camera stays close and handheld, then settles when the key appears.

Shot 1 (0–4s): [Wide, static] Rain streaks the diner window. Marcus and Lena sit opposite each other in a red vinyl booth, two cold coffees between them.
Shot 2 (4–9s): [Medium close on Marcus, handheld] Marcus leans in, jaw tight. Marcus (low, angry): "You told me it was finished."
Shot 3 (9–14s): [Medium close on Lena, handheld] Lena holds his gaze. Lena (calmly): "It was. Until tonight."
Shot 4 (14–18s): [Insert, slow push in] Her hand slides a small brass key across the table and stops beside his cup.
Shot 5 (18–22s): [Two-shot, static] Marcus stares at the key; Lena looks out at the rain.

Requirements: live action, 35mm lens, warm neon from outside mixed with cold fluorescent light, light film grain. Marcus: late 40s, broad build, short grey beard, a worn brown leather jacket over a black T-shirt. Lena: early 30s, slim, dark bob, a mustard raincoat over a cream turtleneck. Their faces and clothes stay the same throughout, no face changes. Marcus has a deep, tired voice; Lena speaks softly with a slight French accent; lips in sync with their lines. Each person appears only once. No background music, no score, no melody: only rain and diner sounds. No subtitles.

Tips

• Dialogue runs at about 2.5 words a second: a 4-second shot fits about ten words. Longer lines get rushed or cut off.
• Put each line in the shot where it is spoken, with the speaker's name, so the right person's mouth moves. Do not quote words of the line again afterwards; repeated words tend to end up as on-screen text.
• Keep each line in one language, and name it if it is not English ("says in Japanese").
• Give each speaker's voice and accent in a few words, and use the same words in every video of a series.
• Write "No subtitles." Without it, the model sometimes writes the dialogue on screen. It cannot be fully prevented: a reference video with text on screen can override it, and portrait videos get subtitles more often than landscape.
• With Sound on, the model may add its own music. To keep it out, name it several ways ("no background music, no score, no melody") and say it at the start and again at the end of the prompt.
• Show feelings through the body: "jaw tight, fists clenched" works better than "furious", and very intense words can make faces look unnatural.
• One camera move per shot, and slower actions: they come out cleaner than fast, busy ones.
• Say what you want rather than what you do not: "the wardrobe stays closed" works better than "never open the wardrobe".
• One or two things happening per shot. A shot asked to do five things will do two or three of them.