Seedance 2.0 Mini video generator with the official model card

What Seedance 2.0 Mini actually is

Seedance 2.0 Mini is the lowest-cost tier inside ByteDance’s Seedance 2.0 generation. It is not a separate generation and not a community rename: ByteDance lists it on its own model service as doubao-seedance-2-0-mini-260615, and carries it as a named model on its own Dreamina product under the label Dreamina Seedance 2.0 Mini.

The one place it is missing is the research site. On the official Seed model index the 2.0 generation appears as a single entry called Seedance 2.0, and there is no separate model page at the mini path. That absence is worth stating plainly, because it is the main reason the pages ranking for this term disagree with each other: half of them treat Mini as an unverified label, and the other half invent specifications for it.

Two dates in the official model IDs make the family history checkable. Seedance 2.0 and Seedance 2.0 Fast both carry the suffix 260128, and Mini carries 260615. On the same official table, a later suffix means a later addition to the line, so Mini joined the family roughly five months after the other two.

Everything below separates three things that are easy to blur: what ByteDance publishes about the model, what other companies publish about their own products built on it, and what this workspace sends when you press generate.

The official model card, line by line

These rows are read from ByteDance’s model list for its own model service, which is the developer-facing surface where a model ID has to exist before anything can call it. The models listed beside it on that page are the reference for every comparison on this page.

Note what the capability row does not say. There is no reduced list, no missing reference mode and no removed editing path. The tier is defined by resolution and price, not by a shorter feature list, and the next section shows that this holds across the whole family.

Three tiers, one official table

Seedance 2.0, 2.0 Fast and 2.0 Mini are published as three rows on the same official model list, so the differences between them are readable from a single source rather than assembled from competing marketing pages.

Read across the capability row and the three tiers are identical. Read across the resolution row and Fast and Mini are identical to each other, while only the standard tier climbs past 720p. So on the official table, the whole family differs along one axis and the middle tier is separated from the cheapest one by nothing that the table states.

That gap is the honest shape of this comparison. The official page does not publish what separates Fast from Mini, and this page does not invent a difference to fill it. The next section gives the one number that does separate them, which the same publisher prints on a different official page.

The official price, tier by tier

ByteDance publishes list prices for all three 2.0 tiers on its own model pricing page. This is the only numerical separation between Fast and Mini that either publisher prints, and it is unusually concrete: the same table gives a per-clip and a per-second figure for each tier at the same settings.

The same page gives a second set of rows for requests that carry a video input, with the price rising as the input clip gets longer. At 480p with a 2 to 15 second input, Mini runs from 1.27 to 2.81 CNY per clip, Fast from 1.99 to 4.42, and the standard tier from 2.53 to 5.62. At 720p the same three run from 2.72 to 6.05, from 4.28 to 9.50, and from 5.44 to 12.10.

Token rates are published as well. Mini is listed at 23.00 CNY per million tokens when the request carries no video and 14.00 when it does, both at 480p and 720p, against 37.00 and 22.00 for Fast and 46.00 and 28.00 for the standard tier at the same resolutions.

A limited-time discount is running on that platform: from 14:00 Beijing time on 7 August 2026 to 14:00 on 7 October 2026, Mini’s 480p and 720p output is billed at 40 percent of list. The official text says that at 720p this lowers the floor to about 0.2 CNY per second, which is consistent with the table above, where 0.50 multiplied by 0.4 gives 0.20. Fast runs a 75 percent discount over the same window, and the standard tier discounts only its 1080p output.

Two boundaries belong with these numbers. They are list prices on ByteDance’s own developer platform, in yuan, for API calls that platform meters by tokens; they are not this site’s prices, and nothing on this page converts them into credits. And a discount window is a time-boxed promotion on that platform, not a property of the model.

What this workspace exposes for this tier

The picker is the authority on what you can submit here, and it is narrower than both the official model card and the official product page. The rows below are read from this site’s own model configuration, so they describe this workspace rather than the model.

Four rows here disagree with the official model card, and they are worth naming rather than smoothing over. This picker offers 1080p for a tier whose official resolution list stops at 720p, and it offers no 480p at all, which is the resolution the official price examples use for the cheapest case. Both are differences between this site and the official platform, so treat the resolution you select as this workspace’s request rather than as a published property of the model, and read the quote before you submit.

The audio row is the third one. Seedance 2.0 and Seedance 2.0 Fast expose a generate-audio switch in this workspace and Mini does not, so nothing here should be read as a claim that the tier cannot produce audio. It is a control this picker does not send.

One more row is worth stating for the same reason. The official duration range for the tier is 4 to 15 seconds, and this picker starts at 2. A 2 or 3 second request is this workspace’s shape, not the card’s.

Mini, Fast and 2.0 inside this workspace

The rows below are this site’s own configuration for the three tiers, which is why they can differ from the official model card in the previous section. They are the differences a visitor here actually experiences.

The one row that matches across all three is the reference ceiling, and it matches the official card’s spirit: all three tiers accept one reference image in image-to-video and one reference clip in reference-to-video through this picker.

The pricing implication is the official table in the section above, not this table. Mini is the cheapest of the three there at every resolution the three share, and Fast sits between it and the standard tier.

How the official guide says to write the prompt

ByteDance publishes a prompt guide for the Seedance 2.0 series on the same developer documentation site as the model list, and it is the only first-party writing advice that covers this tier. It gives a base formula per task type and then a longer formula for the general case.

Three details in that guide change how the prompts are written in practice. First, it says to use shot numbering rather than seconds, because the model is unstable on exact timestamps and forcing them can degrade the result, so a prompt that reads 0-3 seconds / 3-6 seconds is working against the guidance while a prompt that reads Shot 1 / Shot 2 follows it.

Second, it asks for one camera move per shot, on the grounds that stacking a push, a pan and a tilt in the same shot makes the frame less stable.

Third, it reserves four different bracket pairs for four kinds of information, one each for music, sound effects, spoken dialogue and on-screen captions, and it requires dialogue to stay in one language. A small-language line has to be labelled with its language. The guide’s own examples are published in Chinese and this page does not paraphrase them.

Precise subject + action detail + scene and environment + light and colour + camera movement + visual style + image quality + constraints

Reference the <subject N> in <image N>, and generate ...  /  Reference the <motion, camera, style or sound> in <video N>, and generate ...  /  Reference the timbre in <audio N>, and generate ...

Strictly edit <video N>, changing its <original attribute> into <new attribute>. Anything not mentioned stays unchanged.

Extend <video N> forwards or backwards, and generate ...

Keep it subtitle-free. Avoid generating any text or subtitles. Do not generate a logo. Do not generate a watermark.

Shot 1: camera move, subject action and expression, position or spatial change, audio. Shot 2: ... Shot 3: ...

The failure modes the official guide names

The same guide has a troubleshooting section that is more useful than most prompt advice, because it names the failure, gives the cause and gives a fix. These are its headings, in its order.

• Character identity drift. The published symptom is a generated face that does not match the reference, or that changes mid-clip. The guide blames face references that are too small a share of the image or that are mixed in with full-body and costume references, and its fix is to add a dedicated head-only reference shot, keep the face and costume references separate, and put the most important reference earliest in the prompt. It also warns against multi-view character sheets, because the model may read the angles as different people.
• Unwanted subtitles. The guide states plainly that it cannot be avoided with certainty, and offers three ways to lower the odds: an explicit no-subtitles constraint, removing text from the reference material before generating, and generating in landscape first, because landscape produces subtitles noticeably less often than portrait.
• Logos and watermarks. The fix is the same shape as the subtitle case, an explicit instruction not to generate a logo and not to generate a watermark.
• Style drift. The symptom is a request for a 2D or 3D illustration style that comes back photoreal, usually because the reference image is photoreal and the prompt never names a style. The guide’s fix is to name the style in the prompt, and to convert the reference image into the target style before generating if precision matters.
• A visible jump where an extended clip is stitched to the original. The guide’s answer is post-production rather than prompting: trim six frames off the tail of the first segment and one frame off the head of the second, at every join. It also suggests ending a generated segment on a cut and starting the next one in the new scene.
• Doubled characters. In crowded frames the model can place two copies of the same person, and the guide attributes it to unclear subject definitions and to multi-view reference sheets. Its fixes are to label each character with its reference image in a fixed format, to add a global instruction against identical-looking duplicates, and to avoid using a full script as the prompt.
• Quality loss when extending. Chained extensions accumulate degradation, especially on faces, and the guide’s remedies are to convert the source clip into a plain white 3D model before extending, to prefer the sharpest available reference images, and to limit how many times a clip is extended.

Jobs this tier fits

The capability list from the official table and this picker’s control surface together describe a narrower job than the standard tier does, and the pages competing for this term name a similar set. The list below stays inside what one of those two sources supports.

• Short social cuts at 9:16, 1:1 or 16:9. This picker covers all three, and every clip in the gallery below is between 4 and 15 seconds.
• Product clips built on one product photograph, using image-to-video with a single reference image.
• Variant testing on a fixed shot, where the point is to compare several openings or treatments rather than to finish one of them.
• Reference-guided motion, where a short clip decides the pace and camera rhythm and the prompt supplies new content.
• Drafting before a final render. The official price table is the argument here: Mini is the cheapest of the three tiers at every resolution the three share.
• First-frame animation, since first-frame to video is on the official capability list even though this picker exposes it as image-to-video.
• Short extensions, since video extension is on the capability list and the guide documents how to stitch the join; this picker does not expose an extension control.
• Storyboard and beat drafting with the guide’s shot-numbered structure.

What the other Mini pages claim

Four pages hold this term. Three are third-party generators and one is a page about a third-party generator, and none of them publishes a model ID. The rows below record what each one states and which part of it is first-party.

How to generate here

Four steps, in the order the workspace asks for them.

1. Pick the mode: text to video, image to video when one product or character photo should anchor the clip, or reference to video when a short clip should drive the camera.
2. Write the prompt, or open one from the official formula list above and adapt it. Keep one camera move per shot.
3. Set duration between 2 and 15 seconds and the resolution you need, then attach the reference image in image-to-video or the reference clip in reference-to-video.
4. Read the quote, sign in when the workspace asks, then submit. If a task fails, read its status before retrying so you do not submit a duplicate.

What this page does not promise

A model page should say where its own claims stop. These are the boundaries for this one.

• No free allowance is promised. Whatever credits or trial grant exist are shown in your account, not here.
• The prices in the official table are ByteDance’s list prices in yuan on ByteDance’s own developer platform. They are not this site’s prices, and the discount window is a promotion on that platform.
• 1080p is offered by this picker for a tier whose official resolution list stops at 720p, and 480p is not offered at all. Both are differences between this site and the official platform.
• No audio switch is sent for this family from this workspace. That is a limitation of the picker, not a statement that the tier is silent.
• Mini is not the newest generation. ByteDance lists Seedance 2.5 above it, with a 4-to-30-second range.
• Every clip in the gallery below was published by someone else and downloaded from their page. None of them was generated on this site, and no prompt on this page is a promise that the same clip will come back.
• Where a publisher pairs the same clip file with different text in another language, this page reproduces the text from the page named in the caption. A prompt shown beside a clip is that publisher’s own card copy and is not a verified generation record.
• Latency, price and queue behaviour depend on the service at the moment you submit, not on anything stated here.

Where the numbers and clips come from

Every figure on this page traces back to one of the sources below, and the clips trace back to the two publishers named on their cards.