← Back to the AI Workflow Guide
Model Reference
Seedance 2.0
A practical reference, split into two clearly-labeled halves so you always know what's verified and what isn't. Jump to any section:
A practical reference, split into two clearly-labeled halves so you always know what's verified and what isn't. Jump to any section:
From the Dreamina platform, Segmind API spec, and BytePlus documentation.
From ByteDance's official guide. This is the recommended prompt structure.
From the guide: "Describe the rhythm as if you're talking to an editor." Seedance understands human-like rhythmic descriptions rather than photography jargon.
From ByteDance's official guide. These are the movements the model is trained to understand. Use exactly one per generation.
Moves physically toward the subject. Creates intimacy, reveals detail.
Moves away from the subject. Reveals context, creates distance.
Rotates the camera horizontally. Reveals adjacent information.
Rotates the camera vertically. Dramatic reveals, scale emphasis.
Moves alongside the subject. Following action, walk-and-talks.
Circles around the subject. Product shots, heroic reveals.
High-angle ascending or descending. Landscapes, establishing shots.
Micro-sway, documentary feel. Adds urgency and intimacy.
These three rules are stated directly in ByteDance's official prompt documentation.
The official guide calls this the #1 mistake. Mixing camera motion with character motion in one clause generates chaos.
Directly from the guide: "Fast is the keyword most likely to degrade video quality." If you need speed, pick ONE element — subject, camera, or cuts. Never all three.
The official guide states that lighting descriptions have the biggest impact on video quality among all prompt elements. Specify it — don't leave it to the model.
Seedance's own multimodal input system. This is how you reference uploaded files in your prompts.
| Input Type | What It Does | Syntax | Accepted Formats |
|---|---|---|---|
| Image | Locks character appearance, composition, or style | @Image1 | Standard image formats. Clean, well-lit, simple background. |
| Video | Transfers motion, camera work, pacing | @Video1 | MP4/MOV, H.264/H.265, ≤150MB, ≤15s. Max 3 per generation. |
| Audio | Sets rhythm, provides voice/music reference | @Audio1 | MP3, under 15 seconds. |
| Face ID | Strongest identity lock for face consistency | @Image as Face ID | Clear, well-lit, front-facing portrait. |
These come from multiple independent creators and third-party guides. They're not in the official documentation but have been reported to work consistently.
Multiple guides report that the first 20–30 words of your prompt carry the most weight. Put the subject and their core action first, before camera or style.
Slow, continuous actions consistently produce the highest-quality output. If a scene doesn't require speed, write it slow.
A specific syntax structure for camera instructions that community testers report gives the most consistent results.
Pairing one of these terms with any camera instruction reportedly improves output stability:
To build longer sequences (30s, 45s, 60s+), chain generations — each extends the previous output. Describe ONLY what happens next; don't re-describe existing content.
Use "one continuous take, no cuts, uninterrupted camera movement" to prevent the model from inserting edits between extensions.
When you run out of upload slots, prioritize in this order:
1. Motion references (camera/choreography) → 2. Subject consistency (character/product) → 3. Mood/style
When in doubt, cut the style reference — words handle style better than files do.
The model reportedly doesn't reliably infer what each uploaded reference is for. Always state the role in your prompt.
Failure patterns reported by multiple Seedance 2.0 users. Not in the official documentation, but consistent enough across independent reports to take seriously.
A note on "quality keywords" you'll see elsewhere: Many Seedance guides online list terms like "4K Ultra HD," "no jitter," "stable face without deformation," and "normal human structure" as magic-word constraints. These are commonly repeated across guides that cite each other rather than independent testing, and likely originate from older diffusion model workflows. They may help, but they haven't been verified to change Seedance 2.0's behavior specifically. This page excludes them. What actually controls output quality is structural — the prompt formula, camera syntax, lighting description, and clean reference images.