How to Use Seedance 2.5: a Practical Guide

Using Seedance 2.5 on OiiOii comes down to four moves: describe the shot well, anchor it with references, pick the duration and ratio deliberately, and review the result against what you asked for. Everything else is refinement. This guide covers the full working loop, including what the model will and will not let you control.

All product behaviour described here follows the product team's confirmation dated 14 October 2026.

Prerequisites and setup

Seedance 2.5 runs inside OiiOii for members. There is no separate account, API key or integration to set up, and no local hardware requirement. You sign in, open a project or canvas, and select the model.

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

The working envelope is worth memorising before your first generation, because half of all wasted credits come from assuming a control exists:

Bring a concept and any reference material you hold the rights to use. Text alone is enough to start. Where you have an image of the character, product or location, it will do more work than another paragraph of description.

Step-by-step instructions

Step 1: Decide what the shot has to accomplish

Before writing a prompt, say in one sentence what the viewer should understand after watching. "The courier realises the train is the one she was waiting for." That sentence tells you the action, the camera, and roughly how long the shot needs to be.

For the credit math behind each clip, see our breakdown of Seedance 2.5 pricing.

Shots that fail usually failed here, not in the prompt wording. A shot asked to carry three ideas at once returns a result that carries none of them clearly.

Step 2: Write the five-part prompt

Cover subject, action, camera, setting and mood:

A chef in a navy apron slides a copper pan off the flame and turns toward the pass; handheld camera drifting right, warm tungsten light, late-service fatigue.

Then name what must stay stable. The apron colour, the pan's shape, the direction the light falls. Anything you leave unnamed, the model is free to reinterpret between attempts, and it will.

Visual qualities that you might expect to be sliders belong here too, in words. There are no numeric controls for outline weight, shadow intensity or motion speed, so "a soft rim light and slow, heavy movement" is the correct way to ask, rather than hunting for a parameter panel.

Step 3: Attach and prioritise references

Add up to 50 reference inputs in total, mixing images, video and audio. Then tell the model what each one is for: this is the character's face, this is the room, this is the colour palette, this is the mood of the sound.

Keep the set internally consistent. Contradictory references are worse than fewer references, because the model has to choose, and it will not choose the same way twice. If character identity is the thing you care about most, generate the character as its own asset first and reference that, rather than re-describing them in each shot.

Step 4: Set duration and ratio

Pick a duration between 4 and 30 seconds that matches the arc you described in step 1. Then pick the aspect ratio for where the video is actually going.

Ratio is the one decision with no fix afterwards. Final export does not produce alternative platform framings from a finished video, so a piece that needs both widescreen and vertical is two generation tasks. Decide once, early.

Step 5: Use timestamps when timing matters

If a particular beat has to land at a particular moment, timestamped instructions are supported and are precise. Keep every timestamp inside the duration you selected.

This is different from a tempo map or beat grid, which OiiOii does not provide. You are telling the model when something happens in the prompt, not editing a timeline afterwards.

Step 6: Generate, then watch the whole thing

Resist judging the first two seconds. Watch to the end, then ask:

• Does the action complete, or does it stall and loop?
• Did the camera do what you asked?
• Are the references still recognisable at the end of the shot?
• Does the generated audio belong to this scene?

Step 7: Change one thing

Make a single focused correction and regenerate. Changing the prompt wholesale between attempts means you never learn which element was responsible, and you pay credits for that ignorance repeatedly.

Keep a short log of model, duration, references, the change you made and what it did. Within a dozen generations it becomes a better guide than any general advice, because it is calibrated to your subject matter.

Before writing longer prompts, it helps to know what the model can deliver in one shot. The Seedance 2.5 video generation model guide covers its 4 to 30 second range, 1080p output and reference limits.

What you control, and what you describe

This distinction saves more time than any prompt trick.

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

Controls that exist: duration, aspect ratio, whether audio is generated, which references are attached, local region edits so you can fix part of a shot without regenerating all of it, green screen and white model editing, style filters and shaders, shot order, trimming and compositing at the editing stage, and subtitle text, font, position and appearance timing.

Things you describe in words rather than set: motion speed, outline weight, shadow intensity, lighting direction, camera behaviour, atmosphere, pacing within a shot.

Things the platform does not do: frame rate, frame sampling, keyframes or audio channel settings; frame interpolation or motion presets; beat markers, tempo maps or manual time grids; manual mouth shapes, phonemes or lip-sync keyframes; separate background, character and motion layers composited afterwards; animatic or previs passes; transition editing, colour grading or added effects in the editor; adding score, voiceover or sound effects manually, or adjusting volume and mixing; adding new agents to the team.

Lip sync is automatic and supported without restriction by mode or model, so the correct response to a sync you dislike is to review the result and ask the agent for a revision, not to look for a phoneme editor.

Working past 30 seconds

Thirty seconds is a per-generation ceiling, not a project ceiling. For longer pieces, story anime generates multiple storyboard shots and assembles them automatically.

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

The 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 handle the work between them, Art Director, Scriptwriter, Scene Designer, IP Designer, Character Designer, Storyboard Artist and Sound Director, and the system assigns tasks itself. You can also ask a named agent for a specific revision. There is no configuration step where you build the team.

The leverage point is the storyboard confirmation. Reviewing shots there costs you a read-through; discovering the same problem after the shots are generated and assembled costs you the generations.

You can also import an existing finished video you have permission to use, and have OiiOii infer its script, build a storyboard and redesign characters and scenes. That is an optional starting point, not a required first step.

Editing what you generated

The editing stage supports reordering shots, trimming their duration, and compositing. Local region editing lets you correct part of a shot without regenerating the whole thing, which is the single biggest credit saver in the workflow. Green screen and white model editing are supported.

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

Transitions, colour grading and added effects are not editor functions. A cut that reads as a transition comes from how the shots were described and ordered, not from an effect applied afterwards. If you need a graded, effects-heavy finish, that is a traditional post-production step outside OiiOii, and it is optional rather than part of the product's workflow.

Subtitles are well supported, including automatic lyric subtitles with song alignment, and you can edit subtitle text, font, appearance timing and position. Editing when a subtitle appears is not the same as a beat grid, and the two get confused often.

Common issues and what actually fixes them

The character changes between shots. Generate the character as an asset, reference it everywhere, and name the fixed traits in words. Re-describing from scratch invites drift.

The motion reads as mush. One action, one camera move, confirm it, then add. Complexity added before coherence rarely survives.

The audio is wrong. Describe the intended sound alongside the action and regenerate. There is no track-level mixing stage to repair it in, and no placeholder-then-replace workflow.

The result is framed for the wrong platform. New generation at the right ratio. This is unavoidable, which is why the ratio decision belongs in your planning rather than your export.

You are running out of credits mid-project. Usually the revision allowance was set to zero. Count the regenerations you will realistically need when you plan the shot list, not after.

A worked example: three shots that have to match

The best way to learn the workflow is on a small sequence rather than a single shot, because continuity is where the real technique lives.

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

The brief. A mechanic finds something wrong under the bonnet of a car. Three shots: wide of the forecourt, mid at the engine bay, close on her face as she registers it.

Step one, the character asset. Generate the mechanic on her own first.

A woman in her forties, close-cropped dark hair, navy overalls with the left sleeve rolled past the elbow, a red rag in her back pocket; neutral standing pose, even light, plain background.

Review it. Is the sleeve detail there? Is the rag visible? These are the features that will identify her at a glance in every shot, so it is worth a regeneration here to get them crisp.

Step two, the location asset. Generate the forecourt the same way, as a place rather than as part of a shot.

Step three, the shots. Each prompt references both assets and describes only what changes.

Wide: The mechanic from the reference walks across the forecourt toward an open bonnet, wiping her hands; camera static and low, flat overcast light, unhurried. 8 seconds.

Mid: The mechanic leans into the engine bay and pulls a hose free, frowning; camera over her shoulder, cool shadow under the bonnet, concentration. 7 seconds.

Close: The mechanic straightens and looks off-camera left, the frown settling into something more serious; camera static at eye level, same overcast light, dawning concern. 6 seconds.

Note what each prompt does not do: it does not re-describe her hair or her overalls. The asset carries that. Re-describing invites drift, and across three shots drift is visible.

Step four, the review. Watch the three in order. Does she read as the same person? Does the light match across all three? Does the emotional line move from unhurried to concerned?

Step five, the fix. Suppose the close shot has the right performance but the rag has vanished from her pocket. That is a local region edit, not a regeneration. Fix the region, keep the performance.

Reading a result honestly

A habit worth building: separate three kinds of problem before you react.

Prompt problems show up as the model doing something plausible that you did not ask for. The fix is more specific language.

Asset problems show up as inconsistency across shots. The fix is a tighter asset and fewer contradictory references.

Plan problems show up as shots that are individually fine and collectively meaningless. The fix is upstream, in the story or storyboard, and no amount of regeneration addresses it.

Most wasted credits come from treating a plan problem as a prompt problem and regenerating a shot that was never going to help.

Frequently asked questions

How long can one generation be?
Four to 30 seconds, your choice. Longer pieces are multi-shot projects through story anime.

Can I set the frame rate or add keyframes?
No. There are no frame rate, frame sampling, keyframe or audio channel settings, no interpolation and no motion presets. Movement is directed through description.

How do I control when something happens in a shot?
Timestamped instructions, which are supported and precise, within the shot's chosen duration.

Is there a beat grid for cutting to music?
No. There are no beat markers, four-or-eight-bar cut tools, tempo maps or manual time grids. Subtitle appearance timing is separately editable, which is a different control.

Can I composite layers afterwards?
No. Background, character and motion do not come out as separate elements, and there is no animatic or previs pass. The storyboard confirmation is the planning stage.

Can I add a transition or grade the footage?
Not in the editor, which handles shot order, trimming and compositing. A cut that reads as a transition comes from how the shots were described and ordered.

Can I fix one part of a shot without regenerating it?
Yes. Local region editing does exactly that, and green screen and white model editing are supported. This is the biggest cost control in the workflow.

Can I use a style filter?
Yes. Style filters and shaders are supported.

Can I start from a video I already made?
Yes, if you have permission to use it. OiiOii can infer its script, generate a storyboard and redesign characters and scenes. It is an optional path.

Who checks that my characters do not infringe anything?
You do. The IP Designer designs IP characters; it does not perform infringement checks.

Next steps and official resources

Generate one short shot today and keep the prompt. Then change exactly one element and generate again, so you learn what that element does on your material. That loop, repeated a dozen times, teaches more than any checklist.

For prompt structure in depth, see Seedance 2.5 prompts. For a complete production walked end to end, see the Seedance 2.5 review. For plan and credit terms, see the pricing page, and for the platform overview, 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