Bytedance Seedance 2.5 Video Generation Model Guide

Seedance 2.5 is ByteDance's video generation model, and the practical question for most people is not how it was built but what it will and will not do for a project. On OiiOii it generates shots of 4 to 30 seconds at 1080p, with generated audio, from a described brief and up to 50 reference inputs. Everything beyond a single shot is handled by the story workflow around it.

This guide is a model-level reference: the capabilities, the boundaries, and how each one changes the way you plan. Product behaviour follows the OiiOii team's confirmation dated 14 October 2026.

Access

Photo by Kelly Sikkema on Unsplash

For the bigger picture, read our overview of AI video generation with ByteDance Seedance 2.5.

The model is available to OiiOii members. Sign in, open a project or canvas, select Seedance 2.5. There is no separate ByteDance or Seedance account, no API key, no integration and no deployment step.

It was announced in June 2026 and rolled out through July. As of 14 October 2026 it is fully available, so an absent model is an account or membership question, not a rollout one. Free platform rewards such as signup and check-in credits do not by themselves constitute access, and reward amounts follow campaign rules that change.

Model capabilities

Duration. Each generation runs 4 to 30 seconds, selected by you. Thirty seconds is the ceiling for a single continuous generation.

Resolution. 1080p output currently on OiiOii. There is no higher-resolution tier, and no plan unlocks one.

Bit depth. 10-bit, built in. It is a property of the output, not a switch, so there is nothing to enable and nothing to check.

Audio. Generated by default, and it can be turned off. This is generation, not a mixing stage.

References. Up to 50 inputs in total, combining images, video and audio. The ceiling is combined, not per type.

Aspect ratio. Selected before generation. Export does not create alternative platform framings from a finished video.

Timing. Timestamped instructions are supported and precise, within the selected duration.

Style. Style filters and shaders are supported.

Editing. Local region editing can correct part of a shot without regenerating the whole thing. Green screen and white model editing are supported.

Lip sync. Automatic, with no restriction by mode or model.

Subtitles. Fully supported, including automatic lyric subtitles aligned to a song, with editable text, font, appearance timing and position.

Model boundaries

These are not missing features awaiting a release; they are outside how this workflow is designed. Knowing them is the fastest way to stop wasting generations.

If you are comparing platforms, here is where you can use Seedance 2.5 today.

New to the model? This walkthrough covers how to access Seedance 2.5 step by step.

• No frame rate, frame sampling, keyframe or audio channel settings.
• No frame interpolation and no motion presets.
• No numeric parameters for outline weight, shadow intensity or motion speed. These are described in the prompt instead.
• No beat markers, tempo maps or manual time grids. Timestamp prompting and subtitle timing are the available forms of timing control, and they are different things.
• No manual mouth shapes, phonemes or lip-sync keyframes.
• No separate background, character and motion layers to composite afterwards.
• No animatic or previs pass. The storyboard confirmation is the planning stage.
• No transition editing, colour grading or added effects in the editor.
• No manually added score, voiceover or sound effects, and no volume or mixing control. No placeholder-audio-then-final-audio workflow.
• No multi-format export from a finished video.
• No adding new agents to the team.

How the model reads a brief

Describe subject, action, camera, setting and mood, then name the details that must stay stable.

You can try the ByteDance Seedance 2.5 video generation model directly on OiiOii.

A dock worker in an orange vest hauls a rope taut and glances up at a crane swinging overhead; camera low and tilting up, flat grey morning light, strained concentration.

The model holds what you name and improvises the rest, differently on each attempt. Most apparent unreliability is a prompt that left a decision open.

Match the described arc to the selected duration. A single gesture written for 30 seconds produces a shot that idles; a three-stage action compressed into 5 seconds produces one where nothing reads.

How references behave

A small, internally consistent reference set outperforms a large contradictory one. Contradictions force the model to choose, and it will not choose the same way twice.

Our step-by-step guide explains how to make a music video with AI tools in 2026.

Tell the model what each reference establishes: a face, a silhouette, a logo, a lighting direction, the mood of a sound. For identity across multiple shots, generate the character as an asset first and reference that asset, rather than re-describing the person each time. This is the mechanism by which sequences stay consistent, and it is more reliable than any wording.

Beyond a single generation

For longer pieces, story anime generates multiple storyboard shots and assembles them automatically. A 60 to 120 second project is a multi-shot production.

The story workflow is: idea or script, confirm the script, develop characters and scenes, confirm the storyboard, generate the shot videos, generate voice and music, then edit and export.

Seven agents cover the work: Art Director, Scriptwriter, Scene Designer, IP Designer, Character Designer, Storyboard Artist and Sound Director. Tasks are assigned by the system, and you can address a named agent for a revision. There is no role configuration step, and the roster is fixed.

Characters and scenes are generated separately and then referenced to produce the shot videos. The IP Designer designs IP characters; it does not perform infringement checks. OiiOii emphasises original, brand-safe output and character and logo consistency, and reviewing your reference material and its usage permissions remains your responsibility.

An existing finished video you have rights to can also be imported, with OiiOii inferring its script, generating a storyboard and redesigning characters and scenes. That is an optional starting point.

Cost behaviour

Credits are deducted on task complexity. The platform names four levers: number of storyboard shots, image generation, model type and video duration.

For planning, that means a project's cost is dominated by shot count and the design work those shots reference, not by the length of any single generation. Add a realistic revision allowance; projects run dry during revisions far more often than during first generation. The pricing guide turns the plans into a cost per hundred credits, and the pricing page has the current figures.

Planning implications

Three boundaries change how you plan before you write a single prompt.

1080p is the output. Check it against your delivery spec first. If a 4K master is required, that decides the question.

Aspect ratio is a generation-time decision. Two destinations mean two sets of generations. Discovering this at delivery doubles a budget.

Audio is generated, not mixed. If your process involves a conventional mix, that work sits outside the product, and the generated track is not a placeholder to be replaced later.

Everything else is adaptable. Visual direction that you might expect to set numerically is written in language, and the model responds to it well when the language is specific.

Reading the boundaries as design decisions

A list of absent features looks like a list of shortcomings. It is more useful read as a design position, because that tells you what the product is optimised for and therefore when it will serve you well.

Direction instead of operation. Motion speed, outline weight and shadow intensity are described rather than set. This favours someone who thinks like a director and frustrates someone who thinks like an animator. Neither is wrong; they are different jobs.

Confirmation instead of correction. The story workflow puts review points before the expensive steps: confirm the script, confirm the storyboard, then generate. Compare that to a pipeline where you produce first and fix later. The product is betting that catching problems early beats having powerful repair tools, and on a multi-shot project that bet pays.

Generation instead of assembly. Audio is generated, not mixed. Characters and scenes are assets that shots reference, not layers that get composited. The absence of a mixing stage and of layered output are the same decision applied twice.

Fixed roster instead of configuration. Seven agents, assigned automatically, addressable by name. No setup, and no extension. You trade the ability to build a bespoke team for never having to.

Once you see the pattern, the boundaries stop looking arbitrary, and you can predict them. Anything that is a parameter, a timeline or a track is probably absent. Anything that is a creative instruction is probably available.

A capability checklist for your spec

Run your delivery requirements against this before committing to the workflow, because three of these are hard stops rather than inconveniences.

• Is 1080p acceptable? If not, stop here.
• Does the piece need one aspect ratio, or several? Several means generating several times.
• Is a generated score acceptable, or do you need a mix? A mix is outside the product.
• Do you need frame-level animation control? Not available.
• Do you need layered output for compositing? Not available.
• Do you need a grade and effects pass? Traditional post-production afterwards.
• Do you need lyric subtitles with song alignment? Fully supported and editable.
• Do you need automatic lip sync? Supported, unrestricted by mode or model.
• Do you need characters consistent across many shots? Supported through character assets.
• Do you need to repair part of a shot cheaply? Local region editing, plus green screen and white model editing.

Frequently asked questions

Is this the same model I would get elsewhere?
This guide describes Seedance 2.5 as it behaves on OiiOii, which is where you access it here, with no separate account or API key.

Is it still rolling out?
No. Announced June 2026, rolled out through July, fully available as of 14 October 2026.

What is the maximum single generation length?
Thirty seconds. The minimum is four. Longer pieces run through story anime as multiple shots.

Can I raise the resolution by paying more?
No. 1080p is the current output regardless of plan.

Is 10-bit something I enable?
No, it is built in. There is no toggle and nothing to verify.

Is the 50-reference limit per type?
No, it is a combined total across images, video and audio.

Can I turn generated audio off?
Yes.

How is cost calculated?
By task complexity: number of storyboard shots, image generation, model type and video duration.

What is the most underestimated cost?
The image-generation work that precedes the shots, and the revisions after them.

Can I reuse a character across projects?
Generate the character as an asset and reference it. That is the mechanism for consistency within a project, and the same asset is what you reuse.

Does the storyboard replace previs?
It serves the planning purpose. There is no animatic or previs pass as such.

Can I export a vertical version of a finished widescreen video?
No. Aspect ratio is a generation-time decision.

Is there any timing control at all?
Yes: timestamped prompting within a shot, and editable subtitle appearance timing. There is no beat grid, tempo map or manual time grid.

Who checks IP?
You do. The IP Designer designs IP characters and does not perform infringement checks.

A short glossary

Generation. One Seedance 2.5 task producing one shot of 4 to 30 seconds.

Asset. A character or scene generated on its own and then referenced by shots. The mechanism for consistency.

Anchor. A detail named in the prompt that must not change, such as a colour, a silhouette or a light direction.

Storyboard confirmation. The review point before shots are generated, and the cheapest place to make changes.

Story anime. The path for pieces longer than 30 seconds, generating multiple storyboard shots and assembling them automatically.

Local region editing. Correcting part of a shot without regenerating the whole thing.

Timestamped instruction. A precise in-prompt placement of a beat within the shot's chosen duration. Not a timeline control.

Next steps and official resources

Try one shot, keep the prompt, change one element, try again. That loop calibrates the model to your own subject matter faster than reading about it.

For the working method, see how to use Seedance 2.5. For prompt structure, see Seedance 2.5 prompts. For a complete production reviewed end to end, see the Seedance 2.5 review. For platform context, see the OiiOii workflow overview.

Product details in this guide follow the product team's confirmation dated 14 October 2026; promotional rewards can change.

Related guides

• ByteDance Seedance 2.5 AI video generation
• ByteDance Seedance 2.5 video generation model
• Where can I use Seedance 2.5
• How to access Seedance 2.5
• Seedance 2.5 pricing