Seedance 2.0 Prompting Bible

Seedance 2.0 camera, lighting and motion keywords with a five-layer prompt structure

Tier A · StrongPrompt PackReferenceArticlex.com

A reference article that treats Seedance 2.0 as having its own prompt language and documents it: every camera keyword, every lighting modifier, every constraint the author found to actually work, plus the five-layer prompt structure he uses to build a shot. It states the model's real input ceiling — up to nine reference images (character sheets, mood boards, product photos, storyboard panels), up to three video clips for camera motion, choreography or pacing, and up to three audio tracks — and frames the model as a multimodal film set rather than a text-to-video box. The author says the framework was compiled from hundreds of generations, the official Volcengine documentation, Higgsfield and Yaroflasher tutorials, and confirmed community techniques. Written to be the single tab kept open while generating.

For an agent

This is the one to pull from when a user asks you to write a Seedance 2.0 prompt — it gives you the controlled vocabulary (camera moves, lighting modifiers, negative constraints) and the five-layer ordering, which matters because the model weights layers positionally rather than reading free prose. Copy the input ceilings into any pipeline you build: nine images, three video clips, three audio tracks, each tagged to a role, and fail loudly rather than silently dropping the tenth reference. Two cautions — the keyword list is empirically derived and drifts with model updates, and the article funnels to a paid community for the setup half, so do not promise a user the end-to-end walkthrough is free.

Why it is here. Model-specific prompt vocabularies are usually folklore scattered across threads; this compiles one into a single structured reference with the official docs as a source.

Install command, framework, licence and component list.

Facts

License: Proprietary / all rights reserved
Pricing: Free, Free but account required
Framework: Not applicable
Styling: Not applicable
Distribution: Article, thread, or post
Platform: Social / marketing creative, Web
Content type: Prompts, Prose / writing, Video
Agent readiness: Ships a ready LLM prompt, Login required, Blocks bots
Accessibility: Not applicable
Maturity: Stable
Pricing detail: Free to read with an X account. The full setup walkthrough is gated behind the author's paid community at weeklyaiops.com; the prompting reference itself is not.

Notable for

five-layer prompt structure
camera and lighting keyword vocabulary
documents the 9 image / 3 video / 3 audio input ceiling
cites Volcengine official documentation
setup half gated behind a paid community

Source summary, @EXM7777

398,000 views

Order the prompt subject > action > camera > style > constraints: subject pins a centre of gravity, action is the kinetic anchor, camera locks framing before the model re-decides the lens, style comes late so it does not hijack motion, constraints close the remaining gaps.
Write directions, not states — 'she slowly turns toward the camera, breeze lifting the hem of her skirt' executes as a sequence where 'she looks happy' only gives the model a photograph to approximate.
Always split subject motion from camera motion into two separate clauses; fused phrasing like 'spinning camera around a dancing person' is the actual cause of most output people blame on the model.
Lighting description has the largest single effect on output quality per the official Volcengine guide — if you add only one thing to a weak prompt, add lighting; 'golden hour' is the highest quality-per-word gain.
Describe camera rhythm (slow, gentle, smooth) rather than f-stops, ISO or millimetres, use one primary movement per generation, and sequence compound moves into temporal phases instead of stacking them.
Banned-word list: 'fast' unqualified, bare 'cinematic', 'epic', 'amazing/beautiful/stunning', 'lots of movement', and 'glow/glimmer/glints' (which induce specular flicker) — the rule is that any word describing how the viewer should feel forces the model to guess a visual and it guesses wrong.
Positive constraint phrasing beats negative prompt syntax: 'avoid jitter', 'avoid bent limbs' (in every character prompt), 'avoid identity drift', 'maintain face consistency', plus the standard suffix 'sharp clarity, natural colors, stable picture, no blur, no ghosting, no flickering'.
Untagged reference files get averaged into mush — give every upload an explicit @ role; the first-last frame trick (@Image1 opening frame, @Image2 closing frame) gets coherent interpolated motion with no storyboarding.
Iterate one variable per pass and score for continuity and adherence; rewriting the whole prompt after a failure makes the cause unidentifiable.

(Visit the original source)

Provenance

Verified 2026-08-15. Not fetched per instruction (x.com); record built from the pre-retrieved article text supplied in the batch.

Also in Agent Design Tooling

x.comTier A