LESSON 2 · BASICS

How generation works: the overall process and a map of capabilities

Before diving into the finer points of prompting, it's important to see the whole picture. Seedance 2.0 has a single, predictable pipeline: you pick an input mode, write a prompt, set parameters and evaluate the result. This logic is the same in any interface where the model is available — only the placement of controls changes. Master it once, and you'll work confidently anywhere.

The universal 4-step pipeline

Every generation — from a simple text-driven clip to a complex multimodal scene — goes through the same four stages. Remember them as the skeleton of all your work.

  1. Choose the input mode. First decide what you feed the model: text only, an image as the first frame, a "first + last frame" pair, a set of references (images, videos, audio) or an existing video to edit. The mode determines both what you get and how to write the prompt.
  2. Write the prompt using the formula. Describe your idea with the structure Subject + Action + Environment + Camera + Lighting + Style + Quality Constraints. Remember: the first 20–30 words carry the most weight, one action per shot, verbs in the present tense, clarity matters more than a pile-up of adjectives. If you use references — cite them right in the text with labels.
  3. Set the generation parameters. Choose the model version (Seedance 2.0 for maximum quality or Seedance 2.0 Fast for drafts and tests), resolution (480p / 720p / 1080p, natively up to 4K), duration (roughly 4–15 seconds), aspect ratio (16:9, 9:16, 1:1, 4:3, 3:4, 21:9 or auto) and, if you like, lock a seed for a repeatable result.
  4. Generate and evaluate the result. Watch the finished clip critically: did the motion, composition, light and style land? If something's off — refine the wording in the prompt, change a parameter or the seed and try again. Generation is an iterative process, not a single shot.
✅ Tip: change one thing at a time
To understand what exactly affects the result, on each iteration change either the prompt or one parameter — but not everything at once. That way you'll quickly learn to predict the model's behavior rather than guess.

Map of capabilities: input modes

Seedance 2.0 is a multimodal model: it accepts not only text as input, but also images, video and audio. Below are the five main modes. Choosing the mode is your first and most important choice in the pipeline.

Mode What you feed as input What you get / when to use
Text → video A text prompt only. A completely new video from a description. Maximum freedom: you set the appearance, scene, motion and style with words. The best start when you have no source materials.
Image → video (first frame) One image + a prompt. The uploaded image becomes literally the first frame, and the model brings it to life following your description of the motion. Ideal when you need a precise starting composition or a specific character.
First + last (interpolation) Two images: a start and an end frame + a prompt. The model smoothly builds the transition between the two frames. Use it for controlled transformations, morphing and a precise start and end of the clip.
Multimodal reference Up to 9 images, up to 3 videos (≤15 sec total), up to 3 audio (≤15 sec) + a prompt. The most powerful mode. You take a character from one reference, the scene from another, the style from a third, the rhythm and motion from video/audio. For complex, carefully planned scenes.
Video editing (V2V) A source video + a prompt. The model reworks an existing clip: it changes the costume, background, light, style, while preserving the motion and timing. For when the base clip is already good but needs tweaks.

Universal syntax for referencing references

When you upload several materials, the model needs to understand which one exactly you're referring to at a given point in the prompt. For this there's a single label syntax — it works in any interface with Seedance 2.0. Just drop the label right into the prompt text where you mention the corresponding reference.

Images

Reference with the label [Image 1], [Image 2] and so on — in upload order, up to 9 of them.

Example: "the character from [Image 1] walks…"

Video

Reference with the label [Video 1], [Video 2], [Video 3] — up to 3 clips.

Example: "match the camera motion of [Video 1]"

Audio

Reference with the label [Audio 1], [Audio 2], [Audio 3] — up to 3 tracks.

Example: "sync the rhythm to [Audio 1]"

Here's what a prompt with labels looks like in practice — the model understands exactly which material to apply where:

The woman from [Image 1] walks through the neon street from [Image 2], camera moves like [Video 1], pacing synced to the beat of [Audio 1], cinematic, film grain, moody lighting.
🔑 The main thing: the skill transfers between tools
The same prompting skill works everywhere Seedance 2.0 exists. The prompt formula, the input modes, the label syntax [Image 1] / [Video 1] / [Audio 1] and the generation parameters are the logic of the model itself, not of a specific interface. Only the placement of buttons and fields changes, but what you write and how you think stays the same. Learn the model — and any tool will make sense to you.

In the following lessons we'll break down each step of this pipeline in detail: we'll start with the prompt formula, then dig into input modes, parameters and advanced techniques like multi-shots and the transformation arc.

Try Seedance 2.0 yourself

Generate your first clip in ZenCreator — Seedance 2.0 is available in the Image-to-Video tool.

Start Creating
PreviousNext lesson