How to Write Stable Seedance 2.5 Prompts for 30-Second AI Videos

Camera Language That Stays Stable

Camera instructions work best as a sequence of clear production decisions. Choose one primary move per beat instead of stacking multiple techniques:

• Framing: “wide establishing shot,” “waist-up medium,” “tight close-up,” or “over-the-shoulder.”
• Movement: “static tripod,” “slow dolly-in,” “gentle handheld follow,” “lateral tracking,” or “slow orbit.”
• Screen direction: “she walks left to right,” “camera tracks beside him from his right,” or “end facing the doorway.”
• Focus/composition: “keep her face centered,” “background remains softly out of focus,” “finish on a close-up of the product.”
• Lighting continuity: “warm late-afternoon side light remains constant” is more controllable than changing from daylight to neon night mid-shot.

Avoid using camera jargon decoratively. A “dolly zoom,” Dutch angle, or rapid whip-pan can be useful when it serves the story, but each introduces another constraint that can compete with identity, action, and reference adherence. JXP's broader Seedance camera guide lists techniques including dolly zooms, Dutch angles, and time-lapse—use those as deliberate shot choices, not stacked flourishes.

The 50-Reference System: Budget, Not Target

Seedance 2.5 supports up to 50 uploaded multimodal inputs per generation: up to 30 images, 10 video clips, and 10 audio files. Independent sources report the same allocation and describe using the inputs to carry character, set, palette, motion, and sound constraints.

Treat “50 references” as a budget, not a target. The JXP guide warns that uploading everything you have is the wrong instinct. If you are not filling all 50 slots, prioritize in this order:

1. Identity references: a small, internally consistent set for the character’s face, hair, outfit, and key accessories.
2. World/style references: images that establish the set, color palette, materials, lighting, and product design.
3. Motion/camera references: short clips only when you want a particular performance rhythm, choreography, or camera behavior.
4. Audio references: dialogue/voice, ambience, effects, or musical direction, each labeled by role.

Explicitly label references in the prompt—for example, “Image 1–3: character identity; Image 4: coat; Image 5–7: café interior; Video 1: walking pace; Audio 1: café ambience.” Multiple independent guides recommend assigning references roles so the model can distinguish what each input is intended to control.

Do not use redundant or contradictory references—such as several different faces for one protagonist, incompatible costume versions, or both daytime and night versions of the same locked scene—unless the transition is explicitly part of the story.

Reusable 30-Second Prompt Template

Below is a complete template based on the structure and techniques described above. Adapt the specific scene, references, and beats to your project:

The key constraint is that the extra duration should support a fuller arc, not more unrelated events. The 30-second setting adds time—it does not automatically make a prompt capable of supporting more action. Keep the action simple, the references clean, and the ending explicit.