How to Use Seedance 2.0
Describe the subject, action, scene, camera and sound. In multimodal mode, assign each uploaded reference a role with its matching @image, @video or @audio tag.
LumiYingUpdated
Inputs
Use multimodal references for creative guidance, start/end frames for endpoint images, or video extension to continue a clip. These workflows generate new footage.
Describe the subject, action, scene, camera and sound. In multimodal mode, assign each uploaded reference a role with its matching @image, @video or @audio tag.
References
Seedance 2.0, Fast and Mini support up to 9 images, 3 videos and 3 audio files, with a combined maximum of 12 files. Video references total up to 15 seconds; audio references also total up to 15 seconds.
Describe the subject, action, scene, camera and sound. In multimodal mode, assign each uploaded reference a role with its matching @image, @video or @audio tag.
Pin the opening shot
@Image1 as the first frameFor a specified first or last frame, select start/end frame mode and use its image fields. A reference tag in multimodal mode is creative guidance, not a frame-lock control.
Keep a face consistent
@Image2 is the main characterReview identity, clothing, product details, movement and sound across the whole clip. References guide generation; exact preservation is not guaranteed.
Copy camera movement
Reference @Video1's camera tracking and dolly movementStart with one clear camera movement. Add complexity only after checking that the result stays coherent.
Set the mood with music
@Audio1 as background musicSeedance 2.0 supports synchronized audio generation and audio references. Check speech, music, timing and unwanted sounds before using the result.
Transfer artistic style
@Image3 is the art style referenceColor grading, texture, and visual tone get pulled from this image. One reference is enough — stacking multiple style refs muddies the output.
Combine everything
@Image1 as first frame, @Image2 is the main character. Reference @Video1's camera movement. @Audio1 as background music.Each file has a clear job. This is the core idea: be explicit about what each asset does, or the model guesses.
Write Prompts That Actually Work
Describe the subject, action, scene, camera and sound. In multimodal mode, assign each uploaded reference a role with its matching @image, @video or @audio tag.
Putting It Together
“A woman in her 30s, dark hair pulled back, navy linen blazer, turns slowly toward the camera and smiles. Standing on a rooftop terrace at sunset, city skyline behind her. Medium close-up, slow dolly-in. Soft key light from the left, warm rim light, shallow depth of field, film grain.”
Review identity, clothing, product details, movement and sound across the whole clip. References guide generation; exact preservation is not guaranteed.
Pick Your Settings
Choose the resolution your delivery needs and compare the total credit estimate. A suitable Fast or Mini result can be your final clip.
Duration
4–15 seconds. Start at 4–5 seconds while you dial in the look. Longer clips cost more credits and amplify any problems in your prompt.
Aspect Ratio
Choose 16:9, 9:16, 1:1, 4:3, 3:4, 21:9 or adaptive, according to the delivery format.
Model Variant
Standard Seedance 2.0 offers 480p, 720p, 1080p and 4K on LumiYing. Fast and Mini offer 480p and 720p on LumiYing at lower credit rates than standard at the same resolution and reference mode.
Availability
Fast is designed for lower latency, but request complexity and the queue affect completion time. There is no fixed 2× speed or waiting-time guarantee.
Mistakes That Waste Your Credits
You're going to hit these at some point. Knowing them upfront saves you a lot of trial-and-error credits.
Cramming 3+ characters into one scene
What happens: Faces drift, bodies warp, someone grows an extra arm
Fix: Two characters max per scene. Keep clear physical distance between them. If you need a crowd, make them background elements and only describe one or two in detail.
Multiple camera movements in one shot
What happens: Jittery, unstable footage that looks like a broken gimbal
Fix: Start with one clear camera movement. Add complexity only after checking that the result stays coherent.
Writing a 200-word essay as a prompt
What happens: The model picks random details, ignores others, and invents things you didn't ask for
Fix: Describe the subject, action, scene, camera and sound. In multimodal mode, assign each uploaded reference a role with its matching @image, @video or @audio tag.
Not assigning roles to uploaded files
What happens: Reference images get interpreted randomly — your character ref becomes a background, your style ref becomes a character
Fix: Always use @tags with explicit role descriptions. "@Image1 as first frame" beats uploading an image and hoping.
Expecting readable text in the video
What happens: Signs, titles, and subtitles come out garbled or unreadable
Fix: Either skip on-screen text entirely, or make it large, centered, high-contrast, and single-word. Multi-line paragraphs won't work.
Fast hand gestures or detailed finger work
What happens: Extra fingers, fused hands, impossible anatomy
Fix: Slow down all gestures. "Gently raises one hand" works. "Rapidly gestures while counting on fingers" doesn't.
Tips That Actually Move the Needle
Once you've got the basics down, these are the habits that separate decent results from consistently good ones.
Start short, then extend
Generate a 4–5 second clip first. Lock in the look, the motion, and the character. Then use longer durations or extend the clip. Going straight to 15 seconds with an untested prompt is a credit gamble.
Trim your video references
Don't upload a 15-second clip when you only need 3 seconds of camera movement from it. Trim to the exact segment you want the model to reference. Cleaner input = cleaner output.
One style, one finish
Pick a primary style and one supporting aesthetic. "Wes Anderson color palette with film grain" works. "Wes Anderson meets cyberpunk noir with anime influences" doesn't — the model can't resolve conflicting directions.
Choose by the brief
Choose the resolution your delivery needs and compare the total credit estimate. A suitable Fast or Mini result can be your final clip. Fast is designed for lower latency, but request complexity and the queue affect completion time. There is no fixed 2× speed or waiting-time guarantee.
Iterate one variable at a time
Changed the prompt AND swapped a reference image AND adjusted duration? You won't know which change caused the improvement (or regression). Change one thing per generation.
Front-facing headshots for character ref
Character consistency works best with a clear, well-lit, front-facing portrait. Side profiles, group shots, or heavily stylized illustrations give the model less to lock onto.
Workflow
Choose the resolution your delivery needs and compare the total credit estimate. A suitable Fast or Mini result can be your final clip.
Prep your assets
Gather a character headshot, a style ref, maybe a video for camera movement. Trim videos to the exact segment you need.
Write a structured prompt
Describe the subject, action, scene, camera and sound. In multimodal mode, assign each uploaded reference a role with its matching @image, @video or @audio tag.
Choose by the brief
Choose the resolution your delivery needs and compare the total credit estimate. A suitable Fast or Mini result can be your final clip.
Review the result
Review identity, clothing, product details, movement and sound across the whole clip. References guide generation; exact preservation is not guaranteed.
Common Questions
How long does it take to generate a video with Seedance 2.0?
Fast is designed for lower latency, but request complexity and the queue affect completion time. There is no fixed 2× speed or waiting-time guarantee.
Do I need all four modalities to get good results?
No. Text-only works fine for simple scenes. Adding a reference image for the first frame or a character dramatically improves consistency. Audio and video references are optional power tools — use them when you need rhythm sync or specific camera work.
What's the difference between Seedance 2.0 and Seedance 2.0 Fast?
Standard Seedance 2.0 offers 480p, 720p, 1080p and 4K on LumiYing. Fast and Mini offer 480p and 720p on LumiYing at lower credit rates than standard at the same resolution and reference mode. Switching models generates a new clip. Composition, identity and timing can change; the existing result is not upscaled or repaired in place.
Can I upload a video and just change one element?
Use multimodal references for creative guidance, start/end frames for endpoint images, or video extension to continue a clip. These workflows generate new footage. Review identity, clothing, product details, movement and sound across the whole clip. References guide generation; exact preservation is not guaranteed.
My characters keep changing appearance between shots — how do I fix that?
Describe the subject, action, scene, camera and sound. In multimodal mode, assign each uploaded reference a role with its matching @image, @video or @audio tag. Review identity, clothing, product details, movement and sound across the whole clip. References guide generation; exact preservation is not guaranteed.
Is there a way to extend a clip beyond 15 seconds?
A Seedance 2.0 output is 4–15 seconds. To continue a scene, use Video Extender for a separate continuation clip or assemble clips in an editor; inspect the join for continuity.
Creative Inspiration Gallery
Discover stunning AI-generated videos. From photography to architecture, people to creative projects — see how AI transforms ideas into cinematic videos.
Theory's Over. Go Make Something.
You know the @tags, the prompt structure, and the mistakes to avoid. The rest is iteration.