Seedance 2.0 Guide: Inputs and Controls

Seedance 2.0 is ByteDance's reference-led video model for text, image, video, and audio inputs. ByteDance says the upstream model can accept as many as nine images, three video clips, and three audio clips alongside a text instruction, then generate a high-quality 15-second multi-shot audio-video result. Oakgen currently exposes two Seedance 2.0 workflows: text-to-video with optional image, video, and audio references, and image-to-video with an optional last frame.

Those statements are related, but they are not interchangeable. An upstream model capability does not guarantee that every host exposes the same switch. This guide keeps ByteDance's published capability set separate from Oakgen's current controls.

Seedance 2.0 facts checked on August 10, 2026

The last two rows matter. The absence of an audio switch is not proof that every provider response is silent; it means creators cannot currently control that setting in Oakgen. Do not promise a built-in soundtrack, localized clip edit, or extension button solely because ByteDance demonstrated the upstream model. Use only the controls visible in the selected Oakgen workflow and inspect the returned file.

Choose the workflow before writing the prompt

Text-to-video with references

Use Oakgen's Seedance 2 text-to-video workflow when you need several assets to contribute different information. A practical reference packet might contain:

• one character or product identity image;
• one environment image;
• one short motion or camera reference video;
• one audio reference when tempo or sound character matters.

The point is not to fill every slot. Each file needs a declared job. Nine unrelated images give the model nine directions to reconcile.

Image-to-video with an optional last frame

Use the Seedance 2 image-to-video workflow when the first composition already looks right. Add a last frame only when the ending needs a defined destination, such as a closed product, a final pose, or a matched transition.

If you are unsure, use the first-frame, last-frame, and reference decision table.

The controlled reference packet

Write a one-line role for each asset before upload. If two files both own the wardrobe, remove one or decide which wins.

A Seedance 2.0 prompt structure

The Seedance prompting guide contains broader prompt patterns. For reference-led production, use this narrower structure:

GOAL: [the deliverable and viewer response]
IDENTITY LOCK: [unchanged face, wardrobe, product, label, materials]
REFERENCE ROLES: [image 1 supplies identity; video 1 supplies camera pace; audio 1 supplies rhythm]
TIMELINE:
- 0-5s: [one composition and one action]
- 5-10s: [one transition and one action]
- 10-15s: [payoff and final hold]
CAMERA: [one move per beat]
LIGHT AND COLOR: [one coherent visual world]
CONSTRAINTS: [known failure risks only]

Avoid syntax claims the current Oakgen interface does not document. You can refer to assets by order and role in plain language even if a provider-specific @reference convention changes.

A worked production plan: product-in-use short

Suppose the deliverable is a 15-second vertical clip for a travel mug. Prepare a front three-quarter product image, a creator image holding the same mug, and a five-second phone video that shows the desired handheld pace.

GOAL: 15-second vertical product-in-use clip for a commuter audience.
IDENTITY LOCK: Preserve the navy travel mug's tapered body, silver rim, black lid,
handle position, and matte finish. Preserve the creator's face, gray coat, and red scarf.
REFERENCE ROLES: Product image controls mug geometry. Creator image controls identity
and wardrobe. Video reference controls only handheld pace and camera height.
0-5s: Medium view inside a quiet train. Creator places the mug on the window ledge once.
5-10s: Close view. She opens the lid and takes one sip while the camera moves slightly closer.
10-15s: She closes the lid and lifts the mug; hold the product clearly for the final two seconds.
LIGHT: Cool morning window light with a soft warm carriage reflection.
CONSTRAINTS: No label text, duplicate mug, changing lid, new jewelry, camera orbit, or extra hand action.

This is a production plan, not a claimed test result. The prompt reduces ambiguity, but you still need to inspect geometry, hand contact, continuity, and final-frame space.

Seedance 2.0 quality-control pass

Score the result without rewarding surface polish:

When identity or product shape changes, use the Seedance 2 consistency workflow. For broader failures, open the Seedance prompting guide.

Where Seedance 2.0 fits

Seedance is a sensible choice when the job benefits from several reference types or a 15-second internal timeline. A simpler image-to-video model may be easier for a single five-second movement. Another model may suit dialogue generation or a control that Seedance does not expose in Oakgen.

Pick the workflow by job, not by release date. The Seedance, Kling, and Veo creator comparison covers broader model selection; the Seedance 2 versus WAN 2.7 motion-control guide focuses on control methods.

Common Seedance mistakes

Treating every reference as general inspiration. Name the role of each one. Otherwise, the model must decide which asset owns identity, style, set, and movement.

Assuming reference audio and generated audio are the same control. Oakgen currently exposes audio-reference input but no generated-audio switch. Inspect the returned file and plan a separate sound pass when the mix must be controlled.

Starting at 15 seconds because it is available. Use a shorter duration for one action. Fifteen seconds helps when the concept genuinely contains several beats.

Changing identity and camera at once. Large viewpoint changes reveal unseen parts of a person or product. Give the model those views as references or reduce the move.

Rendering approved text. Add claims, prices, disclaimers, and exact logos during editing.

FAQ

What inputs does Seedance 2.0 accept?

ByteDance describes text, image, video, and audio input. Oakgen's text-to-video form currently exposes all four; its image-to-video form uses a required opening image and optional last image.

How many references can I upload in Oakgen?

The current Seedance text-to-video form allows as many as nine images, three videos, and three audio files. It also displays file-size and combined-duration limits beside those fields.

Can Seedance 2.0 generate a 15-second video?

Yes. ByteDance's launch article names 15-second multi-shot output, and Oakgen currently offers duration selections from four through fifteen seconds.

Does Oakgen's Seedance 2.0 workflow generate sound?

The current forms do not expose a generated-audio switch. The text-to-video workflow accepts audio references, while upstream Fal documentation describes synchronized audio. Because Oakgen does not offer a direct toggle, inspect the returned file and finish sound separately when you need an approved mix.

Does Seedance 2.0 support a last frame?

Oakgen's image-to-video form currently exposes an optional last-frame image. The text-to-video form does not expose an end-frame field.

Is 4K available?

Oakgen currently displays a 4K resolution choice for both Seedance 2.0 forms. Check the cost shown in the interface before submitting because pricing can change.

How do I keep a character or product consistent?

Use a sharp identity reference, write one stable identity block, assign roles to other references, limit unseen rotation, and change one variable between retries. Follow the Seedance 2 consistency checklist.

Continue this workflow

• Prepare product video sources and prompts
• Draft, render, and inspect Seedance 2.0 with Oakgen's 4K option
• Build a 15-second Seedance shot list
• Repair character or product drift
• Debug and improve a Seedance prompt

Sources

• ByteDance: Seedance 2.0 official launch, February 12, 2026
• ByteDance: Seedance 2.0 model overview
• Fal: Seedance 2.0 reference-to-video endpoint