Skip to main content

 

Seedance 2.5 is a third-party video model in Runway, developed by ByteDance as the successor to Seedance 2.0. It generates a single video up to 30 seconds long from a text prompt plus up to 50 references, and it can edit or extend video you already have.

Seedance 2.5 is available in Custom mode, Workflows and Agent. This article covers how to access the model, the four creation modes, prompt and reference best practices, available settings, and what to expect from your generations.


Spec information

Supported inputs

Text, image, video, audio

Duration

4s to 30s, or Auto

Aspect ratio

Auto, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16

Output resolutions

480p, 720p

Output formats

MP4, MOV

Maximum references

Up to 50 per generation: 30 images, 10 videos (30s max each) and 10 audio clips

Availability

Custom mode, Workflows and Agent

Credit cost

720p: 30 credits per sec (optional: +15 credits per input video sec)

480p: 20 credits per sec (optional: +10 credits per input video sec)


 

Step 1 — Accessing Seedance 2.5

Seedance 2.5 is available in Custom mode, Workflows and Agent. How you reach the model depends on where you're working.

  • Custom mode — search for Seedance 2.5 beneath the What do you want to create? bar, or select Video and then select Seedance 2.5 from the model dropdown
  • Workflows — add the Seedance 2.5 node to your workflow and connect your inputs to it
  • Agent — name Seedance 2.5 in your prompt, or describe the shot you want and let Agent select the model for you

This guide walks through Seedance 2.5 in Custom mode, where every control is exposed directly. The model itself is the same across all three surfaces, though the inputs and settings available to you differ by surface.

 

Choosing the creation mode

In Custom mode, Seedance 2.5 offers four creation modes:

  • Reference — combine images, video and audio in a single generation, with each input playing a specific role. Use when you want granular control over what gets pulled from each input.
  • Keyframe — set a First Video Frame and a Last Video Frame, then prompt the motion between them. Use when you need precise control over how a shot begins and ends.
  • Edit — change an existing video from a prompt or a sketch. Use when the shot is close but a detail needs to change.
  • Extend — continue an existing video forwards or backwards from a prompt. Use when you need more footage on either side of a shot you already have.

Edit and Extend both take a source video and match its aspect ratio, so neither mode has an aspect ratio control. Edit also matches the length of the input video and has no duration control.


 

Step 2 — Writing the prompt

Your prompt will vary depending on your creation mode and what you're trying to accomplish. In any creation mode, Runway recommends using positive, unambiguous and outcome-focused language.

 

Using the reference inputs

A single generation accepts up to 50 references: 30 images, 10 videos of up to 30 seconds each and 10 audio clips. In Keyframe mode you can also set a first and last frame alongside your references.

Reference each input directly in your prompt and describe the role it should play.

  • With images, you might use one reference for a subject and another for a background or accessory.
  • With video or audio, you can preserve motion or sound while changing the style, or keep the structure while swapping out other elements entirely.

 

Editing a Video with a Sketch

In Edit and Extend modes you can draw directly on a frame of your source video instead of describing a change in words. The drawing becomes an image reference attached to that moment in the video.

  1. Select Sketch on the video
  2. Scrub the video to the frame you want to change
  3. Draw on the frame
  4. Select Export Sketch

Your annotation is added to the prompt as @Image 1 at ss:SS, pointing the edit at that frame. Sketches work best in Edit mode.


 

Step 3 — Generating the video

Before generating, review your settings:

  • Audio — toggles audio generation on or off
  • Aspect ratio — Auto, 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16
  • Resolution — 480p or 720p
  • Duration — Auto, or a specific length between 4s and 30s
  • Output Format — MP4 or MOV

Which settings appear depends on your creation mode:

Mode Audio Aspect ratio Resolution Duration Output Format
Reference

Yes

Yes

Yes

Yes

Yes

Keyframe

Yes

Yes

Yes

Yes

Yes

Edit

Yes

Matches input

Matches input

Yes

Extend

Yes

Matches input

Yes

Yes

Select the Generate button to start the generation. Total processing time will depend on your inputs and selected output duration and resolution.

 

Troubleshooting

In the event of generation errors, review the details below to learn more about troubleshooting:

Unable to start a generation

If you're unable to start a generation because the button is greyed out, this is due to the references not meeting the model's input requirements.

Hover over the greyed Generate button to learn more about the required adjustments and review the Spec table for more details on input requirements.

 

Next steps

You now have a generated video ready to download or pull into your next project. From here, you might want to:

  • Assemble your clips in Studio — trim and arrange multiple generations into a longer, cohesive video
  • Upscale for delivery — once you are happy with a generated video, select it, and choose Upscale below.

Example generations

Broadway Jazz × K-pop Routine

Prompt

Broadway jazz x Kpop routine using this art direction

Reference images

Magazine cover reference image 1 of 10 defining the art direction Magazine cover reference image 2 of 10 defining the art direction Magazine cover reference image 3 of 10 defining the art direction Magazine cover reference image 4 of 10 defining the art direction Magazine cover reference image 5 of 10 defining the art direction
Magazine cover reference image 6 of 10 defining the art direction Magazine cover reference image 7 of 10 defining the art direction Magazine cover reference image 8 of 10 defining the art direction Magazine cover reference image 9 of 10 defining the art direction Magazine cover reference image 10 of 10 defining the art direction

 

Animal Thieves in Extreme Slow Motion

Prompt

Live action photoreal: three animal thieves in suits flee through Manhattan on a sunny day, chased by animal police. The fox trips, the loot bag bursts open, and the camera orbits the treasure suspended in extreme slow motion before time snaps back and everything crashes down. Use @Image 1 for the fox, @Image 3 for the street and @Image 12 for the police. Bright late morning sun, low contrast, shot on 35mm. Audio: no music, diegetic sound only.

Reference images

Reference image 1, the fox thief in a suit with a leather duffel Reference image 2, the crown Reference image 3, the Manhattan street Reference image 4, the pocket watch
Reference image 5, the pearls Reference image 6, the cash Reference image 7, the ring Reference image 8, the jade statuette
Reference image 9, the jeweled egg Reference image 10, the matryoshka Reference image 11, the engraved gold ingots Reference image 12, the three police officers

 

Cereal Spot

prompt

A kid pours Loopo cereal, dives into the bowl, and each flavour opens into its own world: a golden honey lagoon he surfs on a spoon, a cocoa dune canyon he rolls through inside a giant ring, a pink berry sky he bounces across. Hard cut on every flavour swap, then return to the real kitchen with the three boxes behind the bowl. Use @Image 1 for the product design. Bright, saturated and joyful. Audio: no music, diegetic sound only.

Reference images

Product sheet reference showing the three [BRAND] cereal boxes and their rings

 

Colossus Reveal

Prompt

2.39:1, photoreal. A cloaked knight walks across red tidal rock toward a mist bank, passing hooded figures kneeling on the shore. Keep the framing low so nothing above the mist is visible. The water shivers, birds scatter, the light dims. The camera cranes up with her gaze to reveal a colossal stone-skinned god filling the frame, two orange eyes igniting beneath its brow. She draws a sword and holds its gaze. Use @Image 1 for the knight and the landscape. Hazy backlight, heavy haze. Audio: no music, diegetic sound only.

Reference images

Reference image establishing the cloaked knight, red tidal rock, canyon walls and mist bank

 

Creator Cold Open

Prompt

15 seconds, A creator in her early twenties sits at a ring-lit bedroom desk talking fast at camera, hyping a sealed collectible box: she holds it up, shakes it beside her ear, cracks the lid and snaps it shut, then points down the lens for the title-card cut. One continuous take with a slow push-in. Shot on a phone camera, teal LED mixed with warm fairy lights. Audio: no music, her voice and diegetic sound only.

Reference images

Reference sheet for the creator cold open example

 

Get granular with directions

Full prompt

@Image 1 efines the [BRAND] running shoe: its exact shape, the bone white midsole and the moulded white logo mark on the lateral side. Ignore its black studio background. defines the runner and the visual style only: the man, his face, build, sweat-darkened long sleeve, shorts and stride, and the cold-warm dawn grade. Ignore its environment entirely: ignore the wide multi-lane avenue, its tall buildings, its shuttered shops and its street layout. The setting of this film is described in Shot 5 and is an upscale American suburban neighbourhood, not the street in Image 2. He is already wearing the shoes from Image 1. Use exactly one runner and one pair of shoes. Shot 1: lateral tracking backward along a warm dim apartment at first light, soft golden dawn low across wooden floors. Through a bedroom doorway, held in passing: an unmade bed, duvet thrown back in one motion, pillow still dented. The track keeps moving. End state: the doorway slides away toward the kitchen. Shot 2: the same lateral track carries past the kitchen table in the foreground: an untouched breakfast, coffee full in its cup catching the warm light, a chair left angled. The track does not stop. End state: the table slides away toward the hallway. Shot 3: the track reaches the front door and eases to a hold: coats on hooks, an empty key dish, the cat already sitting in the hallway facing the door, tail curled around its feet. Off screen a phone starts ringing, unanswered into the warm rooms. End state: the cat holds its watch on the door as the second ring sounds. Shot 4: hard cut on the third ring: exterior, extreme close up in pure profile on the shoe from Image 1 mid stride, the white moulded logo mark filling frame, brick sliding out of focus behind it. As the sole meets the asphalt the motion ramps into slow motion for one readable instant, the midsole compressing, grit skipping off the surface, first sun raking the logo mark, then speed releases. Shot 5: cut to the frontal medium shot of the runner from Image 2, camera low ahead of him, travelling backward at his pace: he runs down the middle of a quiet street in an upscale American suburban neighbourhood at dawn, wide manicured lawns on both sides, large colonial and craftsman houses set back behind them, every window dark, white porches, mature trees arching over the street, a sprinkler running alone on one lawn, newspapers lying on driveways, a basketball hoop over a garage, clean parked cars on the driveways, the low sun flaring between the trees behind him. He runs through two full breath cycles, and a small private half smile arrives and settles, the expression of a man winning something no one is watching. The camera eases to a stop as he runs on into the flare. Warm golden dawn interior, honeyed low light, against the cool early street outside with its warm sun flare. Shot on 35mm anamorphic, tungsten-balanced 500 ISO colour negative, photochemical grain, natural halation, physically based lighting, real fabric, brick and asphalt response. No person ever appears inside the apartment and nothing moves there except the cat's tail. The apartment track is lateral and one directional, never reversing. No phone is ever shown, it is heard only. The street stays free of people and moving vehicles, parked cars only, every business closed. The only branding in the film is the logo mark on the shoe. Keep the film entirely free of on-screen text and subtitles. Audio: NO MUSIC. Raw diegetic SFX only

Reference images

Reference image 1, the running shoe defining its shape, midsole and logo mark Reference image 2, the runner defining wardrobe, stride and colour grade

 

FAQ

Should I use Seedance 2.5 or Seedance 2.0?

Both are available in Runway, and each is stronger for different work.

  • Choose Seedance 2.5 when a shot needs to run longer than 15 seconds in a single take, when you are conditioning on audio alone, or when a reference video or audio clip runs past 15 seconds
  • Choose Seedance 2.0 for 1080p output, which Seedance 2.5 does not support
  • Choose Seedance 2.0 Fast or Mini for cheap iteration, since Seedance 2.5 costs more per second and tops out at 720p
Can Seedance 2.5 generate 1080p video?

No. Seedance 2.5 outputs 480p and 720p. For 1080p, use Seedance 2.0.

How many references can I use in one generation?

Up to 50 in total: 30 images, 10 videos of up to 30 seconds each and 10 audio clips. In Keyframe mode you can set a first and last frame in addition to those references.

Can I edit or change a video I already have?

Yes. Edit mode changes an existing video from a prompt or a sketch drawn on a frame, and Extend mode continues a video forwards or backwards. Both accept references, and both match the aspect ratio of the video you drop in.

Why can't I set the aspect ratio or duration when editing a video?

Edit and Extend take their aspect ratio from the video you drop in, and Edit takes its length from that video too, so those controls are locked rather than hidden. Editing may shift the output length very slightly against the source.

Can I use timestamps to control exactly when something happens?

Timestamps guide pacing rather than mark frame-accurate edit points. Use them to allocate roughly how long an event should take or to anchor a single key beat, and expect actions to land close to those boundaries rather than exactly on them. For frame-level timing, edit the generated video in post.