Seedance 2.0 Tips: The Settings That Change the Output

Affiliate disclosure: This article contains affiliate links. If you sign up through one of them, we may earn a commission at no extra cost to you. This never affects our ratings or conclusions.

Most Seedance 2.0 tips you find online are prompt tips. This page is the other half: the settings and input limits that decide whether a prompt ever gets a fair hearing. The short answer is that four numbers do most of the damage — a duration ladder that runs 4 to 15 seconds, reference images that must sit inside an aspect ratio of (0.4, 2.5) and (300, 6000) pixels, reference videos capped at 3 clips / 15 seconds total, and a resolution ceiling that is 720p on Fast and Mini but 4K on the full model. Every number on this page was checked against BytePlus ModelArk’s own documentation on 2026-08-24.

If you want prompt grammar — shot syntax, camera language, copy-paste examples — that lives in our Seedance 2.0 prompt library and the complete Seedance 2.0 guide. This page deliberately does not repeat them.

The four settings that actually change the output

Source: BytePlus ModelArk model capability and parameter documentation, checked 2026-08-24. Everything below explains one row.

The duration ladder: why 4-5 seconds is the right first rung

Seedance 2.0’s duration parameter takes whole seconds in the range [4, 15], or -1 to let the model choose a length inside that range. That is the official range for all three 2.0 variants — Seedance 2.0, 2.0 Fast and 2.0 Mini all list 4-15 seconds (checked 2026-08-24). For context, Seedance 1.0 runs [2, 12], Seedance 1.5 Pro runs [4, 12], and Seedance 2.5 stretches to [4, 30]. The floor moved up at 2.0: you cannot ask Seedance 2.0 for a 2-second clip the way you could on 1.0.

The practical tip is to treat those numbers as a ladder rather than a slider. A first generation is a question about framing, subject identity and camera behaviour — all of which are visible in the first second or two. Generating fifteen seconds to answer that question means paying for thirteen seconds of footage you are about to discard, and video generation is billed by output, not by how much of it you keep. Start at 4 or 5 seconds, confirm the shot is what you meant, then climb.

There is a second reason to climb slowly. BytePlus’s own troubleshooting notes describe character ID drifting — the generated character stops matching the reference partway through a clip, sometimes described as a face swap mid-video. Longer clips give drift more room to happen. A short clip that holds identity is a better base for extension than a long clip that loses it halfway.

-1 is the setting most people never touch. It hands length selection to the model within the supported range, which is useful when a prompt already describes a beat structure and you would rather not fight the model over pacing. It is not useful when you need a clip to land on an exact runtime for an edit.

Reference tagging: name every asset you upload

This is the tip with the most confusing documentation, so here is what is actually verifiable.

BytePlus publishes reference syntax in two places, and the two places do not write it the same way. The Seedance 2.0 prompt guide writes tags inline, attached to the subject they describe — for example Woman in red @Image 1 inside a shot description. The Seedance 2.0 API tutorial writes the same idea in brackets in its worked examples: Use the first-person POV framing from [Video 1] throughout, and use [Audio 1] as the background music throughout... opening frame is [Image 1]. In one of the code samples on that same page the brackets are dropped entirely and the prompt reads opening frame is Image 1. All three forms are official, published by ByteDance, and checked on 2026-08-24.

What the documentation is consistent about is the pattern, not the punctuation. The recommended formulas are stated plainly: for image reference, “Reference <Subject_N> in <Image_N> to generate…”; for video reference, “Reference <Action/Camera_movement/Style/Sound_effect> in <Video_N> to generate…”; for audio reference, “Reference the timbre in <Audio_N> to generate…”. In every case the asset is named in the prompt text and bound to the specific attribute you want taken from it.

That binding is the part people skip, and BytePlus’s troubleshooting section documents what happens when they do. Under duplicated characters, the stated root cause is that “the character subjects are not clearly defined in the prompt, so the model cannot accurately distinguish different character roles” — the result being two identical people in one frame. Under style drifting, the stated cause is a realistic reference image plus a prompt that never named the target style, so the output slides toward live action.

One honest caveat, because it matters: the official docs describe under-specified references as producing confused output. They do not state anywhere that an untagged asset is silently dropped from the generation. That is a common claim in community write-ups and we have not been able to source it.

Two related constraints worth knowing before you plan a shoot. Audio reference on the Seedance 2.0 series is marked as not standalone — it must be used together with an image or a video. And the 2.0 series does not accept direct uploads of reference images or videos containing real human faces; ModelArk routes that use case through separate options such as preset digital characters (checked 2026-08-24).

Reference image limits: the numbers that get an upload rejected

A reference image has to clear four gates before the model ever sees it (BytePlus ModelArk, checked 2026-08-24):

• Aspect ratio (width/height): (0.4, 2.5). A 9:16 vertical frame is 0.5625 and passes. A very tall banner crop at 0.35 does not.
• Width and height: (300, 6000) pixels. Both dimensions, independently. A 280-pixel-tall thumbnail fails even if it is wide.
• File size: under 30 MB per image, and the whole request body must stay under 64 MB. The docs specifically advise against Base64-encoding large files.
• Format: .jpeg, .png, .webp, .bmp, .tiff, .gif — with .heic and .heif additionally supported on Seedance 1.5 Pro and the Seedance 2.0 series. That HEIC support matters if you shoot references on an iPhone and would rather not convert.

Counts depend on the task type. First-frame image-to-video takes exactly 1 image; first-and-last-frame takes 2; omni reference-to-video on the Seedance 2.0 series takes 1-9. Seedance 2.5 raises that last figure to 1-30 — see our tier comparison if you are deciding which version to work in.

The count you should use is much lower than the count you may use. BytePlus recommends 4-5 assets in total — one or two character images, one scene image, one camera-movement video, one audio clip — and adds an explicit warning: “It is not recommended to use the full asset limit. Too many assets will make it difficult for the model to judge feature priorities.” The documented symptoms of overloading are style conflicts, blurry subject identification and drift from intent. There is a matching note for people: past four reference characters, output stability is documented as decreasing, with the wrong number of people appearing in frame. The suggested workaround is to group characters into images of at most four, then use those grouped images as references.

Aspect ratio: the silent crop that causes frame jumps

This is the failure mode most likely to be misread as a model problem. When you run image-to-video and the output ratio you selected does not match the aspect ratio of your uploaded image, ModelArk crops the image — centred — to fit. BytePlus documents the consequence directly: when the input image and the output video disagree on ratio, you get frame-to-frame jumps in the video image.

The documented fix has two steps, and skipping either one leaves the problem in place:

1. Crop the input image yourself to a width and height that match the target output dimensions.
2. Set ratio to adaptive so the model preserves what you gave it instead of imposing its own crop.

If you want to pre-empt that crop rather than discover it, you can work it out yourself. BytePlus states that the crop is centred and points to a separate “image cropping rules” page for the detail, but we were not able to open that page to confirm its exact wording on 2026-08-24, so treat what follows as the standard centre-crop arithmetic you can apply on your own rather than as a quote from BytePlus. Call your image W wide by H high and the target ratio A:B. Compare W/H against A/B. When the original is the taller of the two, the crop is driven by width: it keeps the full width, so Crop_W = W and Crop_H = (B/A) × W, starting at x = 0. When the original is the wider of the two, the crop is driven by height: it keeps the full height, so Crop_H = H and Crop_W = (A/B) × H, starting at y = 0. In both cases the crop is centred on the other axis. Run those two lines before you upload and you will have a close estimate of which part of your frame survives — which is the difference between choosing your composition and discovering it.

The general guidance is softer but useful: keep the specified output ratio as close as possible to the actual aspect ratio of the uploaded image. Since Seedance 2.0 offers six fixed ratios plus adaptive, in practice this means deciding your delivery ratio before you shoot or generate the reference still, not after. If you are working from stills, our image-to-video tutorial covers preparing sources for that.

Reference videos: 3 clips, 2-15 seconds each, 15 seconds total

Video references are governed by their own set of limits on the Seedance 2.0 series (checked 2026-08-24):

The total-pixel-count row is the one that surprises people, because it is a separate gate from the width and height range. BytePlus expresses it as the product of width and height falling between 614×664 and 3326×2494. A 4K master at 3840×2160 is 8,294,400 — inside the ceiling by about six hundred pixels. A 4096-wide DCI frame is not. Downscale masters before you use them as references.

The three-clip ceiling shapes how you plan. Since the totals are shared, three 5-second references exhaust your budget exactly, and a single 15-second reference leaves no room for a second one. Choose what each clip is for — action, camera movement, style, sound — rather than uploading the same scene three times.

Also worth knowing before you build a workflow on repeated extension: BytePlus documents image quality degradation when a generated video is fed back in as an extension input, compounding across multiple continuations, with mottled colour blocks appearing especially in faces. Its own advice is to limit how many times you continue rather than to stack continuations. Separately, jump cuts at extension join points are documented as a known issue whose current recommendation is to fix it in post by aligning keyframes. If your project is a multi-shot sequence, our multi-shot storytelling guide is the companion piece.

There is no draft mode on Seedance 2.0 — build one by hand

A tip that circulates widely is “generate a cheap draft first, then upscale”. On Seedance 2.0 that feature does not exist. In ModelArk’s model capability table, draft mode is marked ✗ for Dreamina Seedance 2.5, Seedance 2.0, 2.0 Fast, 2.0 Mini, and both Seedance 1.0 variants. The only ✓ in that row belongs to Seedance 1.5 Pro (checked 2026-08-24).

What you can still copy is the method. ModelArk describes how draft mode carries a result forward: it automatically reuses the model, content.text, content.image_url, generate_audio, seed, ratio, duration and camera_fixed from the draft when it generates the final video, so key elements stay consistent. That list is effectively a checklist of what you must hold constant to do the same thing manually on 2.0 — minus seed, which the 2.0 series does not accept (see the first caveat below):

1. Generate at 480p or 720p while you are still deciding the shot.
2. Record the exact ratio, duration and audio setting of the take you liked.
3. Re-run with every one of those held constant, changing only resolution.

Two caveats. First, one control on that list you cannot borrow is seed: BytePlus lists seed support for Dreamina Seedance 1.5 Pro, Seedance 1.0 Pro and Seedance 1.0 Pro Fast, and it does not appear in the parameter list for the Seedance 2.0 series (checked 2026-08-24). So if reproducibility by seed is central to your workflow, that is a reason to stay on 1.5 Pro or 1.0 rather than something to configure on 2.0. Holding the remaining inputs constant is the closest equivalent on 2.0, and it is not a guarantee that a 720p and a 4K generation will match shot for shot — we have not verified how closely they track. Second, if you are chasing cost rather than fidelity, switching to Seedance 2.0 Mini or Fast for the exploration pass changes the model, which is exactly the variable draft mode holds fixed; and both cap out at 720p, so they cannot be your final render if you need 1080p or 4K.

If 4K is the goal, budget for the throughput limits rather than just the credits: non-4K generation on Seedance 2.0 allows individual users 180 requests per minute at concurrency 3, while 4K drops to 15 requests per minute at concurrency 1. 4K also comes out in 10-bit H.265/HEVC, which BytePlus warns may not play back directly in some players and browsers. Cost per tier is covered in our Seedance pricing review.

Failure modes, and the setting behind each

All rows above are BytePlus’s own documented symptoms and remedies, checked 2026-08-24, not our test results.

One adjacent note, since the watermark row raises it: a hallucinated logo is a generation artefact, not a disclosure label, and the two get confused. If your finished clips go to a marketplace listing that expects AI-generated media to carry a machine-readable label, DiscloseTag (disclosetag.com) is one tool built for that step — it writes container tags into video files and XMP tags into images, and has a separate check page for re-reading a file to confirm the tag survived (checked 2026-07-28). That is a packaging step after generation and has nothing to do with the settings above.

The pre-generate checklist

Run this before you spend the credits:

• Duration set to 4 or 5 for a first pass, not 15.
• Output ratio matches the reference image ratio, or the image is pre-cropped and ratio is adaptive.
• Every reference image inside (0.4, 2.5) ratio, (300, 6000) px, under 30 MB.
• Total reference videos: ≤3 clips, ≤15 seconds combined, each 2-15 seconds, under 200 MB, 24-60 fps.
• Total assets around 4-5, not the maximum of 9.
• No more than 4 reference characters in one generation.
• Every uploaded asset is named in the prompt and bound to the attribute you want from it.
• Style stated explicitly if the reference is realistic and the output should not be.
• Audio reference paired with an image or video, never alone.
• Resolution actually available on the tier you selected — Fast and Mini stop at 720p.
• Ratio, duration and audio settings of the take you liked recorded, so only resolution changes on the re-run.

Symbol conventions round it out, because they are easy to get wrong and cheap to get right: BytePlus documents parentheses （） for music, angle brackets <> for sound effects, curly braces {} for dialogue and 【】 for subtitles, and asks that dialogue not mix Chinese and English except for proper nouns.

Where this page stops

Everything above is input plumbing. It will not write a shot for you. When the settings are right and the output is still flat, the problem has moved into prompt territory — shot sequencing, camera language, action description — and that is a different craft covered in the Seedance 2.0 prompt library and the complete guide. If you have not picked a platform yet, the access guide covers where each channel is available; note that many third-party front-ends expose only a subset of these controls, so a missing duration or ratio field is usually the interface, not the model.

Want to try the settings on an aggregator that carries Seedance alongside other models? Pollo.ai is one such route, and worth checking against the control list above before you commit to it.

Every number on this page was read from BytePlus ModelArk’s public documentation on 2026-08-24. Model parameters change with releases — re-check the Seedance 2.0 series tutorial and the video generation API reference before relying on a limit for production work.