Seedance 2.0 best settings: resolution, audio and references

For a single Seedance 2.0 shot, start with 5–8 seconds, the final aspect ratio and only the references the shot needs. Use 720p to review the idea if that meets your budget; choose a supported higher resolution when delivery requires it. Enable audio from the first attempt when dialogue or sound timing matters. These are starting points, not settings that guarantee a good result.

The model and provider matter. On BytePlus ModelArk, standard Seedance 2.0 supports resolutions through 4K, while 2.0 Fast and Mini stop at 720p. A host may expose fewer options. This guide covers 2.0, not the newer Seedance 2.5. Model documentation was checked September 10, 2026; the suggested prompts below are worked examples, not newly generated test results.

Seedance 2.0 settings at a glance

Check the model variant before choosing resolution

The current ModelArk Seedance 2.0 tutorial lists these variants:

These are ModelArk’s documented options, not a promise that every Seedance app exposes the same menu. Confirm the model name and generation mode before paying. A missing 1080p option on Fast is not a prompt problem.

Use a lower resolution when the decision is broad motion or composition and the provider’s quote makes that worthwhile. Use the intended delivery resolution earlier when the decision depends on fine product detail, faces or small objects. There is little value in approving a low-resolution preview that cannot show the defect you need to catch.

Seedance 2.0 does not have ModelArk’s dedicated draft mode. A new higher-resolution generation is a new candidate; it should not be treated as the approved clip with extra pixels. If you upscale an existing result instead, inspect the processed file for altered detail. Neither workflow guarantees that a label, logo or face stays correct.

Match the aspect ratio before generating

Choose the frame for the actual placement. A vertical product ad and a landscape website hero need different space around the subject.

For image-to-video, ModelArk documents center cropping when the requested ratio differs from the source image. A wide product photograph can lose the packaging edges when forced into 9:16. Prepare the reference in the intended ratio first, or use the provider’s adaptive ratio option when preserving the original framing is the priority. See the video generation and cropping rules.

Keep critical text and the product away from edges that may be covered by the destination app’s interface. Add exact prices, subtitles and calls to action in your editor so they remain editable and readable.

Choose duration for the action, not the maximum

A 5-second reveal, an 8-second spoken line and a 15-second sequence solve different jobs. Write the action first, then allow enough time for it to happen at the intended pace.

• One visual action: Try 5–8 seconds for a reveal, push-in or simple movement. This is a suggested starting range, not a measured quality threshold.
• Dialogue: Read the line aloud at the intended pace. Include time for the speaker to start and finish naturally.
• Several shots: Label the shots and keep the sequence achievable within the selected duration. Split independently replaceable shots into separate jobs when you need tighter editing control.

For example, a 15-second ad could use three separately generated 5-second shots. At two candidates per shot, you would generate 30 seconds to obtain 15 seconds of selected footage, before any additional revisions. That is a planning example, not an observed acceptance rate. Compare the complete job quote in our Seedance 2.0 pricing guide, including reference inputs and the provider’s billing rules.

Keep audio on when it changes the performance

Seedance 2.0 generates video and audio together, including dialogue, effects and ambient sound. ByteDance describes this in the Seedance 2.0 release.

For a speaking character, test with the intended spoken line from the start. A silent preview cannot establish pronunciation, lip synchronization or whether the sentence fits. Describe the speaker, the exact words and the desired background sound separately.

For a silent product montage with a separately licensed soundtrack, switching generated audio off can simplify the brief. Do not assume this makes the job cheaper or faster: check the selected provider’s current quote. Audio pricing and controls vary by implementation.

When the spoken recording is already approved and the task is to synchronize an existing video, a dedicated lip-sync workflow may fit better than regenerating the scene. Match the tool to the part of the asset you actually need to change.

Use reference roles, not a universal stacking order

There is no documented universal “character → face → style → scene” priority stack. Explain what each supplied asset should contribute, and use the correct input mode.

In ModelArk’s omni-reference prompts, assets are identified by type and number, such as Image 1, Video 1 and Audio 1. Other hosts may use selectable tags. Follow your interface’s syntax and identify each asset explicitly. An asset ID or an unexplained list of uploads is not the same as giving it a clear role.

For strict first-frame matching, use the actual first-frame input. Saying “start with Image 1” in a general reference prompt does not substitute for selecting that mode. Our Seedance reference guide covers image, video and audio tagging in more detail.

Reference limits to check before uploading

ModelArk’s current 2.0 documentation permits up to 9 reference images, 3 reference videos and 3 reference audio clips. Reference videos must fit within 15 seconds combined; audio references have their own 15-second combined limit. Audio-only and text-plus-audio inputs are not supported without an image or video. Check your host for additional combined-file or upload limits.

Real-person portrait inputs also have access requirements. ModelArk directs users to its approved portrait-asset workflows; ordinary uploads containing real faces are not a substitute. A rejected portrait upload may be an access or asset-validation issue, not something to fix by adding more prompt detail.

Three example briefs with settings

The following examples are proposed starting points. They demonstrate how to connect settings to a deliverable; they do not represent tested performance or guaranteed outputs.

1. A vertical product reveal

Settings: Image-to-video, 9:16, 5 seconds, 720p for concept review, audio off if adding a soundtrack later. Use a product image already composed vertically as the first frame.

Prompt: “The matte blue travel mug remains upright on the pale stone counter. The camera moves slowly closer as soft window light passes across its curved surface. Keep the handle, lid and proportions unchanged. One continuous shot; the mug stays fully inside the frame.”

Accept only if: The mug retains its shape, the lid stays attached and the final framing leaves room for the ad copy. If the label is a purchase-critical detail, examine it at delivery size. Add the offer and CTA in editing.

To prepare this kind of asset in Magic Hour, start with image-to-video and select from the models and settings available there. The ModelArk-specific limits above do not describe every Magic Hour model.

2. A short line of dialogue

Settings: An approved character-reference workflow, 16:9, 8 seconds, audio on. Use a character asset accepted by the selected provider.

Prompt: “Medium shot of the illustrated shopkeeper from Image 1 behind a flower stall. The shopkeeper smiles and says, ‘Your order is ready. I added the yellow tulips.’ The camera remains still. Quiet street ambience underneath; no music.”

Accept only if: The sentence is complete, the words are correct, mouth movement follows the speech and the background does not overpower the voice. If the delivery is rushed, shorten the script or allow more time instead of simply raising resolution.

3. A two-shot atmosphere clip

Settings: Text-to-video, 16:9, 10 seconds, audio on, a supported resolution appropriate for review.

Prompt: “Shot 1: A locked wide view of a quiet greenhouse at dawn. Water droplets fall from the leaves onto the tiled floor. Shot 2: Close view of one fern as a drop slides from its tip. Soft dripping and distant birds continue across the cut. No dialogue or music.”

Accept only if: Both requested shots appear, the second image belongs in the same greenhouse and the sound continues naturally. Use your editor for an exact cut time; a timestamp in a prompt is not a frame-accurate edit decision.

Common problems and the next useful change

Change one decision at a time when comparing candidates, and keep the prompt, references and settings with the accepted file. That makes it easier to identify why a result changed and to reproduce the workflow, without assuming deterministic regeneration.

Before publishing the finished video

Watch the full final export with sound. Check identity, product details, unexpected objects, spoken words, cropping and any added text. Review it in the actual placement as well as in a desktop player.

Use the provider’s applicable commercial terms and assets you have permission to use. Download accepted files promptly: ModelArk’s generated video URLs expire after 24 hours. Save the final video and the inputs that produced it, not just the preview link.