Seedance 2.0 Complete Guide: Multimodal Video Creation

Seedance 2.0 is live on WaveSpeedAI. The original Standard and Fast lineup covered 14 endpoints, seven in each tier. The current Seedance 2.0 API catalog also includes Mini routes and lists 21 endpoints in total.

Standard: T2V · I2V · T2V Turbo · I2V Turbo · Video Edit · Video Edit Turbo · Video Extend

Fast: T2V · I2V · T2V Turbo · I2V Turbo · Video Edit · Video Edit Turbo · Video Extend

Seedance 2.0 brings text, images, video, and audio into one creative workflow. A still can set a character or scene; a video can guide motion and camera timing; an audio clip can drive the rhythm. Your prompt tells the model which part of each asset matters. ByteDance’s Seedance 2.0 model card documents the four input types and 4–15-second generation range. The standout feature is this ability to direct different parts of a shot from different references, then refine what the first result gets wrong.

Developers can find endpoint names and integration steps in the Seedance 2.0 API guide. Teams considering a newer version can compare the separate Seedance 2.5 API guide; its limits should not be applied to 2.0.

Quick Specs

Parameter	Specification
Image inputs	Up to 9 reference images
Video inputs	Up to 3 reference videos, 15 seconds combined
Audio inputs	Up to 3 MP3 reference files, 15 seconds combined
Text input	Natural-language prompt
Output duration	4–15 seconds, selected by the user
Audio output	Native sound effects and music when audio generation is enabled
Total file limit	12 files per generation across media types in the original reference-workflow guide; confirm the current endpoint schema
Output resolution	Route-dependent: 480p, 720p, 1080p, or 4K; Turbo routes offer 720p and 1080p

These are limits for the multimodal reference workflow, not a promise that every endpoint accepts every input type. Under the original 12-file limit, the 9-image, 3-video, and 3-audio maxima cannot all be used in one request. Check the chosen endpoint’s schema before preparing a batch.

How to Use References

Give every uploaded asset one clear job. Decide whether you are generating a new clip, changing an existing one, or extending its last frame before you write the prompt.

Entry Points

First/Last Frame Mode: Use Image-to-Video when you have a start image. Its image field is required; last_image can guide the ending.
Universal Reference Mode: Use Text-to-Video with optional reference_images, reference_videos, and reference_audios when several assets guide a new clip. These mode names describe workflows; the API fields vary by endpoint.

Video Edit and Video Extend are separate routes. Both require a source video; Edit changes that footage, while Extend appends a newly generated segment.

The @ Syntax

After uploading files, refer to them by type and upload order. WaveSpeedAI’s current examples use a space: @Image 1, @Video 1, @Audio 1. Say what to take from each file, rather than merely naming it.

Examples of Reference Instructions

Use Case	Prompt Pattern
Set first frame	`@Image1 as the first frame`
Reference motion	`Reference @Video1 for the fighting choreography`
Copy camera work	`Follow @Video1's camera movements and transitions`
Add music/rhythm	`Use @Audio1 for the background music`
Extend a video	`Extend @Video1 by 5 seconds`
Replace character	`Replace the woman in @Video1 with @Image1`

Core Capabilities

1. Enhanced Base Quality

The model is designed to handle connected actions, moving subjects, detailed scene direction, and native audio in one generation. Physics, hands, faces, and small text still deserve a final review; a convincing wide shot can hide a weak close-up.

Example prompt:A girl hangs a piece of laundry, reaches into the basket for the next item, and shakes it out firmly before pinning it to the line. One continuous medium shot. Keep her clothing and the courtyard unchanged. Soft wind, fabric rustle, no music.

The sequence gives the model a clear order of actions. Check whether the fabric remains in her hands and whether the movement carries naturally between beats.

2. Multimodal Reference System

References can guide motion from video, appearance from images, pacing from audio, or a combination of those. A video might be useful for a tracking shot, while an image is the better source for a specific coat or product shape.

Key principle: Name the element to borrow from each file. Use @Video 1 for camera movement only is clearer than use @Video 1 as reference. More files do not automatically give more control; conflicting references can make the target unclear.

3. Character and Object Consistency

Reference images give the model an anchor for a face, product, setting, or color palette across a short sequence. They improve the odds of continuity, but inspect fine details such as logos, signs, and small text before publishing.

Example prompt:The man in @Image 1 comes home tired from work. He slows in the hallway, takes a breath at the door, then reaches for his keys. Inside, his daughter and dog run to greet him. Keep his face, jacket, and the warm interior consistent. Natural dialogue and room sound.

4. Motion and Camera Replication

A reference video can guide choreography, camera direction, or cut timing. Tell the model which aspect matters; an exact frame-for-frame copy is not a safe assumption.

Example prompt:Use @Image 1 for the man’s appearance and @Image 2 for the elevator setting. Use @Video 1 for the slow push-in and orbiting camera move. The doors open, and the camera follows him into the hallway. Keep his expression and clothing consistent across the cut.

5. Creative Template Replication

For an ad or title sequence, a source clip can provide a rough structure: reveal, transition, and final product shot. Replace the source’s subjects with assets you have permission to use, and describe the parts of the template you want to keep.

Example prompt:Use @Video 1 for the reveal timing and camera path. The girl from @Image 1 crouches; wings inspired by @Image 2 open behind her. A wing crosses the lens to transition into an aerial shot. Keep the new character and stone-statue setting consistent; use one continuous camera movement.

6. Video Extension

Video Extend starts from the source video’s last frame and appends a new 4–15-second segment. duration describes the new segment, not the combined video’s final length. An optional last_image can guide where the added segment ends.

Example prompt:From the final frame, the donkey continues riding the motorcycle through the fence. 0–5s: side shot as chickens scatter. 5–10s: close-up of the tire turning in sand, then a high overhead shot of the circular tracks. 10–15s: the donkey launches over a low ridge. Match the source character, light, and camera pace. Finish on the motorcycle moving through a dust cloud.

The Extend route does not expose the same multi-image reference fields as Text-to-Video. Keep the character established in the source clip, or use its supported target last frame when an ending image matters.

7. Video Editing

Video Edit changes an existing clip. The input video is required; a source longer than 15 seconds is trimmed to the first 15 seconds. If duration is omitted, the route detects it from the input. State both the requested change and what should remain untouched.

Example prompt:Replace the red car with a black motorcycle from 2s onward. Match its speed to the original vehicle. Keep the road, camera movement, weather, and other traffic unchanged.

If the original soundtrack must remain intact, set generate_audio to false; the Edit route then preserves the input audio track. For a scene that needs a different plot or new choreography, generate a new shot instead of relying on a small edit to rewrite the source motion.

8. Audio-Synchronized Generation

Seedance 2.0 can generate synchronized sound with the video; supported reference routes can also use audio input to guide mood or rhythm. Name the sound you want, and say whether music, dialogue, or effects should be absent.

Example prompt:Fixed fisheye shot looking down through a circular opening. Use @Video 1 for the lens effect and the horse’s upward glance. Use @Audio 1 for the background rhythm. Add a soft hoof sound when the horse steps forward. No subtitles.

9. Beat-Synced Editing

Audio or video references can guide cut timing for a music-driven piece. Specify which moments should land on a beat; leave room for the outfit or scene change to finish before the next cut.

Example prompt:The girl in the poster changes outfits on the main beats of @Audio 1. The looks come from @Image 1 and @Image 2; she holds the bag from @Image 3. Use @Video 1 only for the pacing of cuts. Keep her face consistent, and shift the light between shots without changing the poster layout.

For a longer montage, @Image 1 through @Image 7 can define successive looks while the music guides the rhythm. Check the total file count before uploading.

10. One-Take Continuity

Ask for a continuous shot when the action needs to feel uninterrupted. Within Seedance 2.0’s 4–15-second output range, keep the path simple enough to fit the time available.

Example prompt:Use @Image 1 through @Image 5 for the runner, stairwell, corridor, rooftop, and final city view. One tracking shot follows the runner upstairs, through the corridor, and onto the roof. End overhead. No cuts; keep the runner’s red coat and the camera direction consistent.

For the spy-thriller variation, set @Image 1 as the woman’s appearance, @Image 2 as the corner, @Image 3 as the masked watcher, and @Image 4 as the mansion. The camera follows the woman in red past the corner and into the mansion without a cut. Review the output for unwanted scene jumps.

Creative Applications

Advertising and E-commerce

Use approved product images to hold a package shape or color while the prompt directs a reveal, lifestyle shot, and soundtrack. Check fine print and logos in the output; reference images do not guarantee exact typography.

Content Localization

Video Edit can change dialogue or soundtrack while retaining much of the source performance. Specify the target language, whether subtitles should appear, and which original sounds to keep. Review lip movements and the final words manually.

Storyboarding to Video

Start with a still in Image-to-Video, or attach multiple panels as references for a new sequence. Give each beat a time range so the model knows the order, then check transitions between panels.

Template-Based Creation

Use a reference video’s camera path or edit rhythm with your own character and setting assets. If the goal is to alter the uploaded footage itself, choose Video Edit; if it is only a guide for a new clip, use a reference-capable generation route.

Best Practices

Be explicit about references. Assign one job to each file: @Image 1 for character identity, @Video 1 for camera movement, @Audio 1 for rhythm.
Prioritize uploads. The original reference-workflow guide specifies a 12-file total cap. Confirm it in the live schema; under that cap, 9 images plus 3 videos leaves no slot for audio.
Check your @ mentions. Number each media type in upload order. A swapped image number can change the character or location you intended to hold steady.
Specify edit versus reference. Video Edit takes the footage to be changed in its required video field. A video attached to Text-to-Video guides a new clip; Video Extend continues from a source video’s last frame.
Align duration and cost. For Extend, request the length of the added segment. Edit can bill input plus output duration; reference videos can also affect billing on Text-to-Video. Inspect the live estimate before submitting.
Use natural language and review the result. Describe the shot as you would to an editor, then check the first and last frames, identity, timing, and audio. Revise the smallest failing part of the prompt before adding more references.

Try It Now

Seedance 2.0 is available on WaveSpeedAI. Pick a route based on the material you have:

New shot from text or a still: Standard T2V / I2V for the full resolution range.
Speed-focused drafts: Fast T2V / I2V, or a Turbo route when 720p or 1080p output fits the brief.
Existing footage: Video Edit to change it; Video Extend to add the next segment.

The original launch announcement advertised a 10% discount through May 13 (UTC+0). That dated offer should not be used as a current price promise. Check the estimate in the playground and the final prediction charge for the route and settings you choose.

What’s Next

Seedance 2.5 now has its own endpoints and input limits. This guide stays focused on 2.0; check the live schema when switching versions or adding a new reference combination.

Need a different model family for a particular look, motion style, or budget? Compare WAN 2.7, Veo 3.1 Fast T2V / I2V, Kling, and Vidu on WaveSpeedAI.

Try Seedance 2.0 Mini if its current speed, output settings, and live price suit the job. New to the family? Start with the Seedance 2.0 API guide.

Previous posts: