Treat Seedance 2.5's reported up-to-50 multimodal input capacity as a filing system, not a target. Give every image, video, audio file, storyboard, and written instruction one role. Rank sources by authority, remove conflicts, then reference stable IDs in the prompt.
Seedance 2.5 is coming soon on Oakgen. Oakgen has not verified the 50-item ceiling, accepted formats, per-modality limits, file sizes, duration limits, role controls, or how providers count a “reference.” Every Seedance 2.5 limit and control below is therefore reported or recommended, not an Oakgen product promise.
Prepare the Reference Pack Before Launch
Review the coming-soon model page, watch the feature examples, and keep a clean reference manifest ready for Oakgen integration testing.
The reference manifest
Do not upload files named final2.jpg, clip-new.mp4, and voice-use-this.wav. Build a manifest that says what each source controls.
| Field | Purpose | Example |
|---|---|---|
| Reference ID | Stable prompt handle | CHAR-FRONT-01 |
| Modality | Image, video, audio, text, storyboard | Image |
| Role | One production job | Character face authority |
| Priority | Conflict order | P0, highest |
| Controls | Properties to copy | Face shape, eye color, haircut |
| Must not control | Properties to ignore | Lighting, wardrobe, background |
| Authority | Owned source, approved design, style-only | Approved character master |
| Time range | Useful segment in a clip | 00:02.1–00:04.8 |
| Status | Approved, draft, rejected | Approved |
This manifest is the reusable asset. It makes conflict review possible before a generation consumes time or credits.
REFERENCE MANIFEST ENTRY
ID: CHAR-FRONT-01 FILE: mara-neutral-front.png MODALITY: image ROLE: character identity master PRIORITY: P0 CONTROLS: face proportions, freckles, eye color, haircut, age range MUST NOT CONTROL: wardrobe, lighting, expression, environment AUTHORITY: approved character master STATUS: approved
Use one role per reference
A beautiful editorial frame may contain the right person, wrong coat, wrong color grade, and an unrelated room. If the prompt says only “use this as a reference,” every visible detail competes for attention.
Assign one of these roles:
| Role label | What it may control | Recommended source |
|---|---|---|
IDENTITY | Face, body, product geometry, prop design | Clean approved views |
WARDROBE | Garment, material, fit, accessories | Front and back design views |
ENVIRONMENT | Layout, architecture, set dressing | Wide establishing reference |
LIGHTING | Direction, softness, contrast, color | A frame with simple composition |
STYLE | Texture, rendering treatment, palette | Style board without key characters |
MOTION | Subject action and timing | Trimmed action clip |
CAMERA | Camera path, framing change, lens behavior | Stabilized camera-reference clip |
AUDIO | Rhythm, ambience, effects, performance | Clean, isolated audio source |
STORYBOARD | Shot order and composition | Numbered frames or script board |
If one file needs two roles, split it into two manifest entries and state which property applies in each use. The file can stay the same; the instruction cannot remain vague.
Priority: P0, P1, P2
Use three priority levels only because the order needs to be easy to audit.
- P0, identity and factual authority: owned product pack shots, approved character masters, exact prop designs, final script or dialogue.
- P1, scene direction: environment, wardrobe, camera move, blocking, lighting, and approved sound timing.
- P2, taste and inspiration: color boards, material samples, edit rhythm, genre, and style references.
P2 must never overwrite P0. If a style frame shows the wrong bottle shape, state that it controls palette and texture only. Better yet, crop the bottle out.
Build six small packs
The reported ceiling may accommodate a large stack, but most jobs need a smaller set. Build packs separately, then combine only the sources required for the shot.
Character pack
Start with a neutral front view, a three-quarter view, a profile, and one full-body proportion reference. Add a scene-specific expression or pose only when the action needs it.
Keep age, face geometry, hair, skin details, height, and any identity marker fixed. Wardrobe belongs in its own pack unless the clothing defines the character.
Use the deeper Seedance 2.5 character consistency guide for a character bible and shot-to-shot checks.
Product and prop pack
Use front, three-quarter, side, and functional-detail views. Record which image proves shape, logo placement, cap, controls, material, dimensions, and moving parts.
Generated text and logos can drift. Plan to composite exact packaging artwork or interface text in post when it must be accurate. Do not use an earlier generated product as a P0 source.
Environment pack
One wide layout reference should define spatial relationships. Add a material or lighting reference only when it contributes something the layout frame cannot show.
Watch for contradictions: morning light in one frame, nightclub lighting in another; door on the left in one plan, on the right in the second. Pick the authority before generation.
Motion and camera pack
Separate subject motion from camera motion. A reference clip with a running actor and a handheld camera cannot tell the model which movement matters unless the prompt says so.
Trim each clip to the useful action. Record start and end times in the manifest.
MOTION-02 controls the runner's three-step acceleration and arm rhythm from 00:01.4 to 00:03.9. Ignore its camera shake, wardrobe, face, and set.
CAMERA-01 controls a slow waist-height lateral track followed by a short push-in from 00:00.8 to 00:04.2. Ignore its actor, action, and color grade.
Audio pack
Split audio into dialogue, physical effects, ambience, and music or rhythm. Use clean files when possible. A reference containing all four lanes gives the model less room to understand which feature matters.
For dialogue, provide approved words and pronunciation separately from performance direction. For music, state whether the source controls tempo, instrumentation, edit points, or emotional movement. Do not use a commercial song unless you have the required rights.
The Seedance 2.5 audio, lip-sync, and editing guide covers the planning work in more detail. Its controls also remain reported until Oakgen integration testing.
Storyboard and text pack
Number the shots and keep one line per beat. Each storyboard frame should define composition and staging, not silently introduce new wardrobe, character, or product designs.
SHOT 01 / 0–4s Composition: refer to BOARD-01 Character: CHAR-FRONT-01 and CHAR-FULL-01 Environment: ENV-LOFT-01 Motion: subject turns toward window; no walking Camera: locked medium-wide Audio: city ambience from AMB-CITY-01 at low level Must preserve: face, jacket, product on desk, window position
A reference budget by job
These counts are production recommendations, not Seedance 2.5 limits.
| Job | Lean starting pack | Add only if the first test needs it |
|---|---|---|
| Character close-up | 3 identity images, 1 expression, 1 lighting frame | Profile or voice performance reference |
| Product demonstration | 4 product views, 1 environment, 1 motion clip | Functional-detail image or camera clip |
| Dialogue scene | 4 identity views per lead, set frame, approved script, clean voice references | Wardrobe backs, blocking, ambience |
| Music-led social clip | Style board, storyboard, motion clip, rhythm or music reference | Character or product pack |
| Short narrative | Character pack, environment pack, numbered storyboard, camera references, audio lanes | Shot-specific props and expressions |
Do not fill unused capacity. Every added file is another potential disagreement.
Conflict audit before generation
Run this check across the manifest:
- Does more than one P0 reference control the same property?
- Does any lower-priority frame contradict product or character identity?
- Does a motion clip contain camera movement that should be ignored?
- Does an environment reference introduce the wrong time of day or layout?
- Do audio files compete for dialogue, rhythm, or ambience?
- Does a storyboard introduce new props or wardrobe without an authority source?
- Is any generated output being treated as factual source material?
- Are rights or consent unknown for any input?
Resolve a conflict by removing a file, cropping it, downgrading its authority, or narrowing its role. Do not write “prefer image 1 if there is a conflict” and then keep ten avoidable conflicts in the stack.
Prompt structure for a large reference pack
The prompt should read like a routing table.
GOAL Create a 15-second product-led scene with three numbered beats.
P0 INVARIANTS
- PRODUCT-FRONT-01 and PRODUCT-SIDE-01 control exact bottle geometry, cap, material, label placement, and color.
- CHAR-FRONT-01 and CHAR-FULL-01 control character identity and body proportions.
- SCRIPT-01 controls exact spoken words. Do not add claims.
P1 SCENE DIRECTION
- ENV-KITCHEN-01 controls room layout only.
- LIGHT-MORNING-01 controls soft left-window lighting only.
- MOTION-POUR-01 controls the hand and pour action from 00:02–00:04.
- CAMERA-TRACK-01 controls the lateral camera path only.
P2 TASTE
- STYLE-BOARD-01 controls warm film texture and restrained palette.
TIMELINE 0–4s: [beat] 4–10s: [beat] 10–15s: [beat]
DO NOT CHANGE Product geometry, label placement, cap, character identity, wardrobe, room layout, approved dialogue, or final product position.
DO NOT GENERATE Extra text, a new logo, unsupported results, duplicate products, extra people, or music not requested in the audio lane.
For scene and camera phrasing, use the Seedance 2.5 prompt guide. The manifest controls sources; the prompt controls the shot.
QA every output by reference role
Review the completed clip against the manifest, one lane at a time.
| Lane | Pass condition | Common failure |
|---|---|---|
| Character | Face, body proportions, hair, age, and wardrobe stay within approved references | Identity blends between two character sources |
| Product | Geometry, color, materials, cap, label placement, and function remain stable | Style frame overwrites package shape |
| Environment | Layout and anchor objects stay fixed through camera movement | Room rearranges between shots |
| Motion | Action timing and physical cause remain readable | Camera reference changes subject movement |
| Camera | Framing and path follow the stated clip or storyboard | Motion becomes an impossible orbit |
| Audio | Dialogue, effects, ambience, and rhythm follow their assigned lanes | Music masks speech or extra dialogue appears |
| Claims and text | No unsupported claim or invented readable text appears | Generated label or on-screen copy looks authoritative |
Use PASS, FAIL, or CANNOT VERIFY. “Looks good” does not tell the next revision what to preserve.
If a lane fails, remove conflicting references before adding more. More input is rarely the first fix.
What Oakgen will verify at launch
Oakgen should not convert upstream reporting into a product guarantee. When Seedance 2.5 integration is live, the model page and generation interface need to confirm:
- total input count and how the provider counts each item;
- accepted text, image, video, and audio formats;
- per-modality count, size, duration, and resolution limits;
- whether role labels, priority, timestamps, or negative controls are exposed;
- reference ordering and duplicate handling;
- video duration, resolution, audio, editing, and credit options;
- upload behavior, failed-job policy, and delivered asset format.
Until those checks pass, use this guide to organize sources, not to promise a specific Oakgen control panel.
Use Seedance 2.0 for work that must ship now
ByteDance's official Seedance 2.0 launch documents combined text, image, video, and audio input, including stated per-type limits for its upstream product. That confirms the Seedance family already uses multimodal reference workflows, but it does not confirm Seedance 2.5's reported 50-item implementation or Oakgen controls.
If your project cannot wait, use Seedance 2.0 on Oakgen and adapt the same manifest to the inputs the current interface exposes.
Common mistakes
Treating 50 as a target
Capacity does not equal clarity. Start lean.
Giving every file equal authority
A product pack shot must outrank a mood board that happens to show a different bottle.
Mixing camera and subject movement
Split the roles or state which movement to ignore.
Using old generated frames as masters
Errors compound. Keep owned or approved source material at P0.
Claiming Oakgen controls before launch
The integration is not live. Formats, limits, and tools remain unverified.
FAQ
How many references should I start with?
Use the smallest set that defines every required identity, product, environment, motion, camera, and audio lane. A product shot may need six sources; a narrative may need more.
Can one reference control several properties?
It can, but write separate role instructions. If the unwanted properties are likely to conflict, crop or replace the source.
Should reference order matter?
The Seedance 2.5 ordering behavior is not verified on Oakgen. Use stable IDs and explicit priority so the pack remains readable regardless of upload order.
Can I use a video for both motion and camera?
Only if both are wanted. Otherwise state the role and the properties to ignore, or create separate trimmed references.
Does the reported 50 include the text prompt?
Oakgen has not verified how the provider counts text, files, storyboard pages, or grouped media. Treat 50 as a reported ceiling, not a planning entitlement.
Can I upload copyrighted music as a rhythm reference?
Use only material you have permission to upload and use. A technical input slot does not grant rights.
Will more character images improve consistency?
Only when they add non-conflicting identity information. Redundant images with different hair, age, styling, or proportions can make direction worse.
Is Seedance 2.5 live on Oakgen?
No. Seedance 2.5 is coming soon on Oakgen. The model page carries current access status.
Sources and related guides
- ByteDance Seed: Seedance 2.0 official launch
- Volcengine Developer Community: reported Seedance 2.5 capabilities
- Seedance 2.5 complete guide
- Seedance 2.5 character consistency
- Seedance 2.5 audio, lip-sync, and editing
Keep the Pack Ready for Oakgen Testing
Build the manifest now, follow the coming-soon model page, and compare the first live Oakgen controls against the reported reference workflow.
