Seedance 2.0 Prompt Guide: How to Write Multimodal AI Video Prompts That Work

If you want AI video that nails camera movement and lets you reuse a real character, environment and sound clip in the same shot, Seedance prompts are worth learning properly. ByteDance’s Seedance pushed multimodal references into the mainstream: you can hand it an image for identity, a video for motion and an audio file for rhythm, all in one prompt. Used well, it produces some of the most controllable AI video available.

The trade-off is that Seedance rewards a different prompting style than Veo or Kling. It is built around motion and reference tags, not just description. Write a Seedance prompt the way you would write a Veo prompt and you leave most of its capability untouched. This guide covers what Seedance 2.0 is good at, what changed with Seedance 2.5, the prompt structure it responds to, copy-and-paste examples, and the mistakes that wreck results.

Update, September 2026: ByteDance released Seedance 2.5 on July 31, 2026. It raises single-generation length to 30 seconds and accepts far more reference files. Everything in this guide applies to both versions; the Seedance 2.5 section covers the differences.

What Seedance 2.0 Is Actually Good At

Seedance 2.0 is ByteDance’s video generation model, launched in February 2026. You can reach it through Dreamina (ByteDance’s international creative platform), Jimeng in China, CapCut in supported markets, and several third-party APIs. A few strengths set it apart:

Multimodal references in a single generation. You can upload up to 12 files at once (9 images, 3 videos and 3 audio clips) and tag them so the model knows which file controls what. One image becomes your character, one video supplies the camera path, one audio clip sets the rhythm.
Synchronized audio in one pass. Dialogue, sound effects, ambience and music are generated with the video rather than layered in afterward.
Strong, expressive motion. Seedance treats movement as a first-class part of the prompt. It handles speed, weight and physics well when you describe them directly.
Flexible duration and framing. Clips run up to 15 seconds across six aspect ratios, including 16:9, 9:16 and 1:1. Resolution options depend on the platform, from 480p up to 2K on the API.

Seedance 2.0 also ranks in the top five of the Artificial Analysis text-to-video arena (with audio) as of September 2026, ahead of Kling 3.0 and Veo 3.1.

The practical takeaway: if your project depends on character consistency, a camera move you have already seen somewhere, or matching the energy of an existing track, Seedance gives you levers text-only prompting does not.

Where you can use it (and where you can’t)

Seedance’s global rollout was paused in early 2026 after legal threats from Hollywood studios over copyrighted characters. When Seedance 2.0 came to CapCut on March 26, 2026, the first markets were Brazil, Indonesia, Malaysia, Mexico, the Philippines, Thailand and Vietnam, and the United States was not on the list. CapCut’s version blocks generating from images that contain real faces, blocks unauthorized IP, and adds an invisible watermark. If you are in the US, expect to reach Seedance through third-party platforms instead.

What’s New in Seedance 2.5

ByteDance’s official Seedance 2.5 announcement lists these changes:

30 seconds per generation, up from 15, with multi-round extensions for longer videos.
Up to 30 images, 10 video clips and 10 audio clips as references in one pass (BytePlus describes this as up to 50 multimodal references including text).
Timestamp-level editing: change a specific moment, character or element without regenerating the whole clip.
New reference types, including clay renders (rough 3D blockouts) and motion references.

Seedance 2.5 is rolling out on Jimeng and Doubao in China, to Dreamina subscribers in Europe, Asia, the Middle East and South America, and through the BytePlus ModelArk API (which is not offered in the US).

What this means for your prompts:

Plan in beats. A 30-second clip needs a story with a beginning, middle and end, not one action stretched thin. Write it as three or four short beats in order.
Give every reference a job. With up to 50 files, unassigned references create confusion. Name each one and say what it controls.
Check the tag format on your platform. ByteDance’s own 2.5 examples write tags with a space (“@Image 1”, “@Video 1”). Use whatever format your interface inserts when you click a file.

The Seedance Prompt Structure

Seedance prompts work best when they lead with motion and use a clear three-part backbone:

Subject and movement: who or what is in the frame, and how it moves.
Background and movement: what changes in the environment around the subject.
Camera and movement: how the camera frames and follows the action.

All three parts include movement. That is the core difference. Seedance is less interested in a static catalog of visual details and more interested in what is happening and how. Adverbs that define speed and force (slowly, sharply, wildly, gently) carry real weight.

A plain text prompt following that backbone:

A lone surfer paddles hard against an incoming swell, then rises smoothly to her feet as the wave curls behind her. The wall of water rears up and crashes in slow, heavy spray. The camera tracks low along the water surface, then pushes in tight on her focused expression as she carves down the face.

Adding Reference Tags

Once you bring in references, the structure gets a second layer. Seedance lets you tag uploaded files and give each one a job:

@Image1 for identity: a character’s face, a product, a specific look.
@Video1 for motion and camera movement: the model follows the motion path and applies your visuals to it.
@Audio1 for rhythm, voice or atmosphere.

A good habit is to put your highest-priority reference first. If the character is what must stay consistent, lead with @Image1.

Use @Image1 for the man's appearance and place him in the office setting from @Image2. Replicate the camera movement and facial expressions from @Video1: a slow dolly zoom on his face as he realizes the news, followed by an orbit around the desk. Match the pacing to @Audio1.

This is where Seedance pulls ahead. You are not describing a camera move from scratch and hoping the model gets it. You are pointing at a clip that already does the move and saying “like that.”

Example Seedance Prompts

Each of these is ready to adapt.

1. Text-only, motion-forward

A black stallion gallops across an open salt flat at dawn, hooves kicking up fine white dust that hangs in the air behind it. Thin clouds drift fast across a pink sky. The camera tracks alongside at full speed, then drops back and rises into a high wide shot as the horse shrinks against the vast empty plain. Hooves thundering, wind rushing past.

2. Product loop

A pair of matte black headphones rotates slowly against a deep charcoal background, light sweeping across the metal hinges as they turn. Faint dust particles drift through the beam. Static centered frame with a slow continuous 360-degree orbit, perfect loop. Low ambient hum building into a single clean bass note.

3. Character consistency with reference

@Image1 as the chef, working in the kitchen from @Image2. She plates a dish with quick, precise movements, then looks up and smiles. Apply the handheld tracking motion from @Video1 as the camera follows her hands to her face. Kitchen ambience and the sizzle of the pan from @Audio1.

4. Motion transfer from a reference clip

@Image1 as the dancer in the empty warehouse from @Image2. Replicate the full choreography and camera tracking from @Video1 exactly: the spin, the drop, the rapid pull-back. Keep the dancer's outfit and the warehouse lighting consistent throughout. Sync every movement to the beat in @Audio1.

5. Mood and atmosphere

Rain falls hard on a neon-lit alley at night, water sheeting off awnings and pooling on the asphalt. A figure in a long coat walks slowly away from the camera, reflections rippling beneath each step. The camera holds low and still, then drifts forward at the figure's pace. Heavy rain, distant traffic, a low synth drone.

6. Found-footage short (text-only)

Dashcam footage from a car stopped at a red light on a quiet suburban road at dusk, wide-angle lens with a timestamp in the corner. A goose waddles into the crosswalk, stops in front of the car, and slowly turns to stare straight into the windshield. A second goose joins it. Static dashboard view, slightly washed-out colors. Engine idling, indicator ticking, a loud honk from the goose, and the driver laughing, "Are you serious?"

7. A 30-second story for Seedance 2.5

Beat 1: A small boy in a red raincoat sets a paper boat into a flooded gutter and watches it drift away. Low tracking shot at water level.
Beat 2: The boat races through a storm drain tunnel lit by shafts of light from the grates above, spinning past floating leaves. The camera follows close behind the boat.
Beat 3: It shoots out into a wide canal at sunset where hundreds of other paper boats are already floating. The camera rises into a high wide shot.
Rain on pavement, rushing water echoing in the tunnel, then calm water and distant city sounds. Gentle piano enters in the final beat.

In each one, movement leads and the camera has a clear job. That is what Seedance is tuned for.

Camera, Style and Motion Tips

Lead with motion, not still details. “A flag” gives the model nothing. “A tattered flag snapping violently in gale-force wind” gives it everything.
Use one strong camera move per shot. A single dolly, orbit or tracking shot reads cleanly. Three camera moves in a 5-second clip produce motion soup.
Match your references to your text. Do not ask for “fast motion” in the text while @Video1 is a slow-motion clip. Contradictions show up as flicker and warping.
Put priority references first. Your most important reference belongs at the front of the prompt.
Direct the audio. Seedance generates synchronized sound, so name the ambient layer, any dialogue and the music feel. Silence is a choice too, so say so if you want it.
Give longer clips a progression. For anything past 5 seconds, describe a beginning, middle and end so the model has an arc to follow.

If filmmaking vocabulary is not your first language, the complete AI video prompt engineering guide breaks down the camera and lighting terms worth knowing.

Common Seedance Mistakes

Writing negatives the wrong way

Seedance follows positive instructions best. Negatives mixed into the action, like “the boy can’t stay still,” confuse the model. Flip them: “the boy waves his hands.” Long keyword dumps (“blurry, ugly, deformed, bad hands, low quality…”) do not help either.

What does work is Dreamina’s own recommended pattern: a short closing line or two that names specific exclusions and anything that must stay fixed.

Avoid: text on screen, extra people in the background.
Preserve: the logo on the mug from @Image1.

Keep these lines short and specific, and put them at the end, after the motion, camera and sound.

Overloading static description

Paragraphs of adjectives about texture and color, with no movement, produce a stiff clip. Trim the static detail and spend those words on motion and camera direction.

Contradicting your own references

Mismatched speed, lighting or framing between your text and your uploaded clips is the top cause of flicker. Keep the text and the references pointing the same direction.

Cramming too many subjects

Five focal points split the model’s attention. Pick one primary subject, let the rest be background, and the clip holds together better.

Burying your key reference

Dropping @Image1 halfway through a long prompt makes it easy for the model to under-use it. If consistency matters, that tag goes first.

How Seedance Compares to Veo, Kling and Hailuo

Each major model has a personality, and prompting style should follow it.

Seedance is the reference-and-motion specialist. Image for identity, video for motion, audio for rhythm. Prompt it motion-first.
Veo 3.1 rewards cinematic, natural-language description with dialogue in quotes, plus reference images and first and last frames. See the Veo 3.1 prompt guide.
Kling 3.0 is best when you want up to six labeled shots in one 15-second generation with native audio. See the Kling 3.0 prompt guide.
MiniMax H3 (in the Hailuo app) takes text, image, video and audio references too, and handles big, expressive motion well. See the Hailuo AI prompt generator page.

If you used to write for Sora, note that OpenAI has shut it down; the Sora shutdown alternatives guide covers where to move. For a full side-by-side, read the 2026 AI video generator comparison, and see the shared prompt structure formula that adapts across all of them.

Writing Seedance Prompts Faster

The motion-first structure works, but writing a full prompt with subject movement, background movement, camera move and audio for every variation takes time. Most projects need 10 or 20 prompts, not one.

That is the gap LzyPrompt fills. You describe your idea in plain words and get 10 prompt variations, each with camera style, setting, lighting, action beats, imperfections and sound cues, plus a ready-to-post caption. They are plain text: paste one into Seedance, add your @ reference tags where you need them, and generate. See sample outputs on the Seedance prompt generator page, or try LzyPrompt free for 7 days.

Frequently Asked Questions About Seedance Prompts

How many reference files can a Seedance prompt use?

Seedance 2.0 takes up to 12 files in one generation: 9 images, 3 videos and 3 audio clips. Seedance 2.5 raises that to 30 images, 10 videos and 10 audio clips. Tag each one (@Image1, @Video1, @Audio1) and give it a specific job.

What is the difference between Seedance 2.0 and 2.5?

Seedance 2.5, released July 31, 2026, generates up to 30 seconds per clip instead of 15, accepts many more references, and lets you edit specific timestamps without regenerating everything. The prompting approach (motion first, clear reference roles) is the same.

Does Seedance support negative prompts?

Not as a long keyword list. Dreamina’s guide recommends a short closing “Avoid:” line for specific exclusions (and a “Preserve:” line for anything that must not change), placed after the main description. Everything else should be phrased positively: describe the outcome you want, not the one you are trying to avoid.

Why is my Seedance video flickering or warping?

Usually a contradiction between your text and a reference clip, for example asking for fast motion while @Video1 is slow motion. Align speed, lighting and framing between text and references, and use one clear camera move per shot.

Can I use Seedance in the United States?

Not directly through CapCut or BytePlus at the time of writing. Seedance 2.0’s CapCut rollout and Seedance 2.5’s BytePlus API both exclude the US. US creators typically reach it through third-party platforms that offer the model.

How is prompting Seedance different from prompting Veo or Kling?

Seedance leads with motion and uses reference tags for identity, camera movement and audio. Veo leans on cinematic written description, and Kling on labeled multi-shot structure. Write Seedance prompts motion-first and assign every reference a clear role.

Start Creating Better Seedance Video

What makes Seedance powerful (multimodal references and motion-first design) is also what makes generic prompts fall flat. Lead with movement, give the camera one clear job, tag your references in priority order, and keep your text and your clips in agreement.

Start with any of the examples above, adapt them for your project, or browse the LzyPrompt blog for more AI video workflows.