Seedance 2.0 Reference Video Prompts: 8 Worked Examples

You uploaded a reference clip to Seedance 2.0, typed “copy this,” and got back a video that copied the wrong thing. The camera move you wanted is gone, but the stranger from your reference clip is now standing in your scene. That is the most common Seedance reference problem, and it comes from the prompt, not the model.

Our Seedance 2.0 prompt guide covers the basics: motion-first writing and what @Image1, @Video1, and @Audio1 are for. This post goes one level deeper. It covers the actual upload limits, how to tell Seedance exactly which part of a reference to borrow, the difference between a reference job and an edit job, and eight worked prompts you can adapt today. The same method works on the newer Seedance 2.5, and the last section shows what changes there.

Where Reference Inputs Live in Seedance 2.0

In Dreamina (ByteDance’s creative platform), the reference workflow sits under AI Video, then Seedance 2.0, then the Omni reference mode. That mode is different from the first-frame and first-and-last-frame modes:

First frame mode locks the opening image. The video starts exactly on your picture.
First and last frame mode animates a transition between two images.
Omni reference mode treats your uploads as sources to borrow from. Nothing is locked to frame one. You describe what each file contributes.

When you upload files in Omni reference mode, Dreamina labels them automatically: @Image1 through @Image9, @Video1 through @Video3, and @Audio1 through @Audio3. Numbering restarts for each type, so your first video is @Video1 even if you already uploaded five images.

One warning for API and third-party users: tags are an interface feature, not magic words. On BytePlus ModelArk, references are passed as roles in the request, and typing a filename or bracket command into the prompt attaches nothing. Check how your host labels uploads before copying prompts from a tutorial written for another platform.

The Real Limits (Check These Before You Upload)

These are the per-generation limits reported for Seedance 2.0’s reference mode:

Input	Max count	Length rules
Images	9	Still images, no duration
Videos	3	2 to 15 seconds each, 15 seconds combined
Audio	3	2 to 15 seconds each, 15 seconds combined
Output	1 clip	4 to 15 seconds

Two rules catch people out:

Audio needs a visual partner. You cannot generate from text plus audio alone. At least one image or video reference must be present.
Real human faces get blocked. ByteDance’s face policy rejects uploads containing photorealistic human faces across its consumer apps and its API channels. That includes your own face. Animals, products, illustrated or stylized characters, and footage without visible faces go through far more reliably.

So a doorbell recording of your empty porch works as a reference; a clip of your neighbor at the door may not.

The Four Questions Every Reference Prompt Must Answer

Most failed reference generations miss one of these:

Which file supplies the subject? (“The orange cat from @Image1”)
What happens in the new clip? (the action beats, in order)
Which file supplies the movement or the look, and only that? (“Follow the camera path from @Video1”)
What must not be copied or changed? (“Do not copy the person or background from @Video1”)

Question four is the one people skip. A reference video contains a camera path, a subject, a background, lighting, and a pace. If you don’t say which piece you want, Seedance often takes all of it.

Reference Job vs Edit Job: The Wording Matters

Seedance 2.0 can do two different things with an uploaded video:

Reference: make a brand-new clip that borrows motion, camera, or style from the upload.
Edit or extend: change or continue the uploaded clip itself (replace an object, remove something, add something, or keep the action going).

Creators who use it daily report that the wording decides which job you get. “Reference @Video1” pushes the model toward a new video. “In @Video1, replace…” or “Continue from @Video1…” tells it to modify or extend the original. If an edit keeps coming back as a whole new scene, remove the word “reference.”

8 Worked Seedance 2.0 Reference Prompts

Each example lists the uploads first, then the prompt, then why it works.

1. Borrow only the camera path

Uploads: @Image1 a red vintage scooter on a white background. @Video1 a phone clip of you walking a slow half-circle around a parked car.

The red scooter from @Image1 parked on a quiet cobblestone street at golden hour, warm sunlight on the chrome. Follow the camera path from @Video1: a slow half-orbit at waist height, ending on the front headlight. Do not copy the car, the street, or the lighting from @Video1. Soft street ambience, a distant church bell.

Why it works: it names the one thing to take from the video (the orbit) and lists what to leave behind. Start here.

2. Motion transfer onto an animal

Uploads: @Image1 a stylized cartoon corgi. @Video1 a 6-second clip of a simple dance move (filmed from the neck down, no face).

The corgi from @Image1 performs the dance moves from @Video1 in a sunny backyard, standing on its hind legs, matching the timing of every step and the final spin. Static camera at the corgi's eye level, vertical 9:16 framing. Keep the corgi's proportions and colors from @Image1. Upbeat backyard ambience, paws tapping on a wooden deck.

Why it works: “matching the timing of every step” marks the video as a motion source, and “keep the proportions and colors” stops the corgi from turning into a human dancer.

3. Product with separate lighting reference

Uploads: @Image1 your product photo (a matte black water bottle). @Image2 a photo of a moody studio setup with a single blue side light. @Video1 a slow push-in shot from any product ad you filmed.

The water bottle from @Image1 stands on a dark reflective surface. Use @Image2 for the lighting only: one cool blue side light, deep shadows, no fill. Follow the slow push-in from @Video1, ending on the logo. Do not copy any object or background from @Image2 or @Video1. Condensation beads slowly run down the bottle. Low ambient hum, a single soft click at the end.

Why it works: each file has exactly one job: look, light, and camera move.

4. Doorbell camera framing from your own footage

Uploads: @Video1 a 10-second clip from your real doorbell camera showing your empty porch at night. @Image1 a photo or illustration of a raccoon.

Match the camera angle, fisheye distortion, night-vision look, and timestamp overlay from @Video1. The raccoon from @Image1 climbs onto the porch, sniffs the welcome mat, then drags a pizza box off the step and out of frame. Static doorbell camera, no camera movement. Grainy black-and-white night vision, slight motion blur on fast moves. Include the scrape of cardboard on concrete and faint crickets.

Why it works: the fisheye curve, exposure, and overlay are hard to fake from text, and your reference already has them. For more concepts, see our doorbell camera video prompts.

5. Cut to the beat with an audio reference

Uploads: @Image1 an illustrated skateboarder character. @Video1 any short skate clip filmed from behind (for motion). @Audio1 a 10-second drum loop you have rights to use.

The skateboarder from @Image1 rides down an empty concrete ramp at sunset. Follow the riding motion from @Video1. Every kick push lands on a drum hit from @Audio1, and the ollie at the end happens on the final cymbal crash. Low tracking shot behind the skater, vertical 9:16. Keep @Audio1 as the soundtrack, add wheel roll and a board clack on landing.

Why it works: it ties specific actions to specific moments in the audio instead of saying “sync to the music.”

6. Replace an object in an existing clip (edit job)

Uploads: @Video1 a clip you already generated of a hand placing a mug on a desk. @Image1 your branded mug.

In @Video1, replace the white mug with the mug from @Image1. Keep the same hand movement, camera angle, lighting, and background. Do not change anything else in the frame.

Why it works: no “reference” wording, a clear swap, and “keep everything else.” Keep edit prompts short; every extra sentence invites extra changes.

7. Extend a clip past 15 seconds (extend job)

Uploads: @Video1 your first 12-second generation of a cat stalking a toy mouse. @Image1 the same cat reference you used for the first clip.

Continue from @Video1. The cat from @Image1 pounces, misses the toy mouse, slides across the wooden floor, and crashes softly into a pile of laundry. Same camera position, same afternoon window light, same color grade. Do not introduce new objects or characters; only continue the existing action. Paws skidding on wood, a muffled thump.

Why it works: re-attaching the character image keeps the cat on model, and the “no new objects” line stops a second cat from appearing. Keep the original aspect ratio and plan for two or three segments, since drift builds up.

8. Two animals, one camera move

Uploads: @Image1 a gray tabby cat. @Image2 a golden retriever. @Video1 a handheld phone clip of a slow walk down a hallway.

The gray tabby from @Image1 sits in the middle of a narrow hallway, blocking the golden retriever from @Image2, who approaches slowly. The cat swats once at the dog's nose. The dog freezes, then backs away and lies down. Follow the handheld walking motion from @Video1 for the camera, moving forward and stopping behind the dog. Do not copy the hallway or any person from @Video1. Include a soft hiss and the dog's nails clicking on the floor.

Why it works: each animal has its own image, so they don’t blend, and the handheld reference adds natural phone sway. The consistent characters guide covers keeping recurring pets identical.

How to Prepare Reference Clips

One move per clip. Trim to the single camera move or action you want. A cut in the middle gives Seedance two motions to reconcile.
Match your target framing. For vertical 9:16 output, film the reference vertically.
Keep it faceless. Hands, backs, empty rooms, and objects avoid the face filter and keep strangers out of your result.
Clean motion beats pretty footage. A shaky clip teaches shaky motion.

Troubleshooting Reference Generations

Problem	Likely cause	Fix
The person or object from your reference video shows up in the output	You didn’t say what to leave out	Add “Do not copy the subject or background from @Video1”
The model ignores a reference	The tag appears late or the job is vague	Move the tag earlier and state its job in one sentence
Flicker or warping	Text and reference disagree (fast vs slow, day vs night)	Make the text describe the same speed and lighting as the reference
Upload rejected	Realistic human face detected	Use a faceless clip, an animal, or a stylized character

If flicker persists, the negative prompts guide explains why rewriting exclusions as positive instructions often works better on models like Seedance.

Using the Same Prompts on Seedance 2.5

ByteDance released Seedance 2.5 on July 31, 2026, and it now sits next to 2.0 in Dreamina’s model list (and on hosts like Runway). The reference method is the same, so every prompt in this post works as a starting point. What changes:

Longer clips. Up to 30 seconds in one generation, with the option to extend twice. A story that needed three chained 2.0 clips can fit in one 2.5 generation.
Bigger reference sets. ByteDance’s announcement lists up to 30 images, 10 videos, and 10 audio clips. That makes multi-character scenes (a whole litter of kittens, a band, a cast) practical.
New reference types. Seedance 2.5 accepts a 3D “clay render” or white model as a reference for blocking and camera movement. ByteDance’s own example reads: “Refer to @Clay Render 1 for camera movement, pacing, shot-size transitions, subject trajectory, and blocking. Refer to @Image 2 for character design.”
Time-range direction. You can direct what happens in a specific stretch of the clip, which pairs well with longer outputs.

A 2.5 version of example 8, written for the longer clip:

The gray tabby from @Image1 and the golden retriever from @Image2 in a narrow hallway, handheld phone footage, vertical 9:16. Follow the handheld walking motion from @Video1 for the camera; do not copy the hallway or any person from @Video1. From 0 to 10 seconds: the cat sits in the middle of the hallway as the dog approaches slowly, and the cat swats once at the dog's nose. From 10 to 20 seconds: the dog backs away, lies down, and the cat walks past it with its tail high. From 20 to 30 seconds: the camera follows the cat into the kitchen, where it sits in front of an empty food bowl and stares up at the lens. Soft hiss, nails clicking on the floor, one loud meow at the end.

Check which version your host is running before you plan uploads around the bigger limits. Many third-party hosts still default to 2.0, and a prompt with 12 references will fail on a 2.0 endpoint.

Writing the Text Half Faster

References handle identity, camera, and sound. The text still has to carry the story: setting, action beats in order, imperfections, and sound cues. That is where most people stall.

LzyPrompt writes that text half. Type an idea like “raccoon steals pizza from porch,” optionally pick a camera style such as doorbell camera, and you get 10 prompt variations, each with a scene title, the full prompt, and a ready-to-post caption. The prompts are plain text that works in any model, so you paste one into Seedance and add your @Image1 or @Video1 tags yourself. The Seedance prompt generator page shows how that looks.

FAQ

How many reference files can Seedance 2.0 use at once?

Up to 9 images, 3 videos, and 3 audio clips in one generation. The videos share a combined 15-second limit, and so do the audio clips. Audio only works when at least one image or video reference is also attached.

Can I use a video of myself as a reference?

Usually not. Seedance 2.0 blocks uploads that contain photorealistic human faces, including your own. Faceless footage, animals, products, and stylized characters work.

How do I make a Seedance video longer than 15 seconds?

Generate the first clip, then attach it as @Video1 in a new generation and start the prompt with “Continue from @Video1.” Re-attach your character image each time and keep the aspect ratio the same. Plan for small drift over several extensions.

Do Seedance 2.0 reference prompts work in Seedance 2.5?

Yes, the @ reference approach is the same. Seedance 2.5 raises the limits, allows clips up to 30 seconds, and adds clay render references, so a prompt that works on 2.0 is a safe starting point on 2.5.

Start With One Reference, Then Stack

Start with example 1: one image, one camera reference, one “do not copy” line. Once that behaves, add lighting, audio, and edits one at a time. If you want the scene text, action beats, and captions written for you, try LzyPrompt free for 7 days and spend your Seedance credits on generations instead of rewrites.