tutorials

How to Use Seedance 2.5: Text, Images, References

Oakgen Team4 min read
How to Use Seedance 2.5: Text, Images, References

How to Use Seedance 2.5: Text, Images, References

To use Seedance 2.5 well, start before the model picker. Define the video's job, delivery format, subject invariants, three scene beats, reference roles, camera path, and audio plan. Then generate a controlled first version, review it against the brief, revise only the failed variable, and export the approved result.

Seedance 2.5 is coming soon on Oakgen. Join the waitlist for integration updates. If your project must ship now, use Seedance 2.0.

StageInputOutput
1. DefineObjective, audience, placementOne measurable creative job
2. PrepareCharacter, product, style, motion, audio referencesA labelled reference pack
3. DirectThree beats, camera, soundA generation-ready prompt
4. GenerateModel settings and referencesA first controlled render
5. ReviewIdentity, action, camera, audio, final frameA precise revision list
6. DeliverApproved masterChannel-specific exports

Before access: prepare a production packet

You do not need the model to prepare the highest-leverage inputs:

  • Objective: what must the viewer understand or feel?
  • Placement: 9:16 Reel, 16:9 ad, product page, or previsualization?
  • Scene: what happens at the start, middle, and end?
  • Invariants: face, product label, wardrobe, palette, location.
  • References: which asset controls each invariant?
  • Audio: dialogue, effects, ambience, and music.
  • Acceptance test: what makes the result usable?

A clear acceptance test might be: "The bottle stays identical, the camera makes one smooth push, the label is readable in the final five seconds, and the cap sound lands when the hand closes it."

A multimodal reference workflow example from Oakgen's Seedance 2.5 model page.

Step 1: choose one creative job

Do not begin with "make a great video." Choose:

  • a 30-second product reveal;
  • a character walking through one connected environment;
  • a vertical social hook;
  • a music-led movement sequence;
  • a spatial previsualization test.

The mistake is choosing the model before choosing the job. Seedance 2.5 may offer many controls, but each generation still needs a single dominant purpose.

Step 2: build a labelled reference pack

Start with three to six references:

  1. character or product anchor;
  2. alternate identity or product angle;
  3. environment;
  4. camera motion;
  5. palette or style;
  6. audio direction.

Name files by purpose: character-front.webp, product-label-close.webp, camera-slow-orbit.mp4. A folder full of final-2.png files invites mistakes before generation begins.

Current product material describes up to 50 multimodal references, but the final formats and limits exposed by Oakgen remain subject to provider verification.

Use the smallest sufficient pack

Add a reference only if you can finish this sentence: “This asset controls ___.” If two assets control the same thing differently, choose one.

Step 3: write the scene as a timeline

Use:

  • 0–8 seconds: establish;
  • 8–21 seconds: develop;
  • 21–30 seconds: pay off.

Describe a dominant action and camera move inside each beat. Put the final frame in writing. For a full template, use the Seedance 2.5 prompt guide.

Step 4: select Seedance 2.5 and configure delivery

When the Oakgen integration is live:

  1. open the Oakgen video workspace;
  2. select Seedance 2.5;
  3. choose the supported aspect ratio and duration;
  4. attach the approved references;
  5. assign references in the prompt;
  6. review the cost shown in the interface before generating.

Oakgen has not announced Seedance 2.5 credit pricing. The live interface and launch page will be the sources of truth.

Follow the Seedance 2.5 Oakgen launch

The model page carries the confirmed capability ledger, feature videos, and waitlist.

View the Model Page

Step 5: review the render by category

Do not review by vague impression. Use this scorecard:

CategoryPass questionIf it fails
IdentityIs the same person or product preserved?Strengthen or simplify identity references
ActionDid the intended event happen in order?Rewrite the failed beat only
CameraDid framing and movement follow the brief?Use one dominant move
ContinuityDid lighting and space remain connected?Remove conflicting environment/style inputs
AudioDid sound align with visible action?Shorten cues and identify their timestamps
PayoffIs the final idea readable?Simplify the last beat

Save the score. The second generation should change one or two failed variables, not restart the entire creative direction.

Step 6: use localized editing when the scene is mostly right

Current product material describes region-level revision. Use it for a bounded face, object, label, or background correction when the rest of the scene works.

Do not use local editing to repair a broken story structure or camera path. That is a prompt problem. The audio, lip-sync, and editing guide explains the difference between a local defect and a sequence-level failure.

Step 7: export for the actual channel

Keep an approved master, then derive channel versions:

  • 9:16 with safe space for platform UI;
  • 1:1 or 4:5 for feeds;
  • 16:9 for web and YouTube;
  • clean version without generated captions;
  • sound-on and sound-off review exports.

Output up to 4K does not eliminate composition review. A beautiful wide master can still crop badly into vertical.

What I would test first

Use one product, one hand, one surface, and one camera move. The scene:

The product begins on dark stone. A hand lifts it once. The camera moves from a medium frame into a slow half-orbit. The final five seconds hold on the unchanged front label. Room tone, one contact sound, one closing note.

It is simple enough to score and complex enough to test identity, action, camera, continuity, audio, and payoff.

Continue with the 30-second ads workflow if your goal is marketing, or the character consistency guide for narrative work.

Common mistakes

  • Uploading unlabelled or contradictory references.
  • Writing 30 seconds as one static shot description.
  • Changing prompt, references, duration, and camera at the same time.
  • Using local editing to solve a sequence-level problem.
  • Exporting every aspect ratio from one unreviewed crop.
  • Quoting speculative prices rather than the live Oakgen interface.

Bottom line

The usable Seedance 2.5 workflow is brief first, references second, generation third. Prepare the production packet now, and the launch-day test becomes a controlled evaluation rather than random experimentation.

Get Seedance 2.5 access updates

Join the Oakgen waitlist for priority-access eligibility and launch-pricing information.

Join the Waitlist

Sources and further reading

how to use Seedance 2.5Seedance tutorialAI video workflowimage to videomultimodal references
Share

Related Articles