Seedance 2.0

By Minel Gunesoglu, founder. I build every published score from dated public evidence, arena boards, vendor terms and sample-weighted review records, and I read the community record behind it. No controlled generation run stands behind this page, and none is claimed. Last updated 12 August 2026; re-checked monthly.

Three days after Seedance 2.0 went public on 12 February 2026, ByteDance had letters from the Motion Picture Association, SAG-AFTRA, Disney and Paramount. "By launching a service that operates without meaningful safeguards against infringement, ByteDance is disregarding well-established copyright law," the MPA's Charles Rivkin wrote; Disney called the output a "virtual smash-and-grab of Disney's IP" (TechCrunch, 15 February 2026). ByteDance tightened the face and IP filters in response, and two months later a rival hosting platform reported the fallout: users were calling Seedance 2.0 "dead on arrival", because the restrictions catch the realistic-face work the model launched on (WaveSpeed, 6 April 2026).

No score card holds a story like that. What the card holds is four numbers, and only two of them are the axes they sit on. On the Artificial Analysis Video Arena's text-to-video board, the entry listed as "Dreamina Seedance 2.0 720p" carries an ELO of 1,223 across 18,647 blind votes and sits third, re-read 12 August 2026 (1225 and second on 30 July, so both the figure and the placing moved). Beside it sits a commercial-terms reading taken out of BytePlus's own ModelArk paperwork. The other two rows could not be filled by the evidence their axes call for, and since 19 August 2026 they publish different measurements under different names, which is most of what follows, and the reason the composite fell rather than rose.

What Is Seedance 2.0, and Why the "Seedance AI" Wrapper Sites Aren't It

Seedance 2.0 is a text-to-video and image-to-video model from ByteDance's Seed research team, launched 12 February 2026 (Seed model page). Three surfaces are official and no others are: the consumer app at dreamina.capcut.com, the enterprise API at byteplus.com, and that research page. The model carries its own Wikidata entry, verified 30 July 2026.

Access arrived in stages: China's Jianying app first, then the Dreamina surface, with global consumer access gated behind an invite-only Creative Partner Program and no public waitlist. That was confusing enough that Hacker News commenters were still arguing six days after launch whether the thing had shipped. "Probably because it isn't officially released yet," one wrote in a thread asking why nobody there was discussing it (Hacker News, 18 February 2026). Builders take a separate road through BytePlus, ByteDance's enterprise cloud, which sells the model as an API and addresses three audiences under a heading of its own on that product page, Why creators, developers and businesses choose us (BytePlus, retrieved 12 August 2026). One retraction goes with that: an earlier version of this sentence put a longer description of that audience (businesses, developers, creators who want official API access to a high-end multimodal video generation model), inside quotation marks, sourced to a third-party write-up we hold no capture of, and neither official API access nor who want occurs anywhere in the bytes of the BytePlus page, fetched twice on 12 August 2026. What that page does say also overtakes something this page recorded on 30 July 2026, that only Seedance 1.5 Pro was live on the API: the model list now carries seedance-2-0, seedance-2-0-fast and seedance-2-0-mini, and the product page's own banner reads "Dreamina-seedance-2.0 API Now Fully Available to Clients" (12 August 2026).

Now the part that decides whether a reader ever reaches the real product. A cluster of domains trades on the name: seedance.ai, seedanceai.cc, seedance.io, seedance.tv, seedance-2-ai.com and seedance2.com. None is operated by ByteDance, and none is linked from any of the three official surfaces, checked 30 July 2026. This page therefore reserves the string "Seedance AI" for those sites, calls the product Seedance 2.0 throughout, and uses "Dreamina Seedance 2.0" only where the arena entry or the consumer app requires it. If the name led you to a star rating, check the domain before you trust it.

Seedance 2.0 Clip Length Limit, Resolution and Native Audio Generation

Dreamina's product page, retrieved 30 July 2026, states the working envelope plainly: 720p to 1080p output, clips of 5 to 12 seconds, early access. A wider 4-to-15-second range circulates in third-party write-ups, and the vendor's own figure is the one this page stands behind. Anything longer means chaining clips rather than asking for one long take, and the practical lever on continuity is the mixed image, video and audio reference inputs the model accepts, not prompt phrasing.

Native audio is the headline capability, and the one agency account on file draws a sharp line through it. A creative agency using the model on client work files short-form dialogue under 30 seconds under what works ("the lip sync is convincing for quick product pitches, intros, and CTAs" (Sidekick Studios, 28 April 2026)) and long-form dialogue under what still needs work: "anything over a minute and the lip sync drifts" (same post, same date). Both were mis-set here until this pass: each label and the sentence beneath it ran together inside one pair of quotation marks, and the source separates them with an em-dash. Treat the audio pitch as a social-length feature that does not yet reach narrative runtime. In plain use-case terms, the evidenced sweet spot is short-form advertising and social clips: a vertical product teaser, a Reels or TikTok cut, a talking-head intro, where the vendor's 5-to-12-second ceiling and that under-30-second lip-sync boundary describe the brief rather than fight it. One version caveat and no more: ByteDance announced Seedance 2.5 at Volcano Engine's FORCE event on 23 June 2026 with 30-second clips, 50 reference inputs and native 4K, all enterprise-beta announcement material rather than shipped capability, and nothing here is measured on 2.5.

Seedance 2.0 Review: What the 80.1/100 Score Card Actually Measures

Before the number, the scope, because on this product the scope is the number's biggest caveat. This card scores Seedance 2.0 as delivered on BytePlus ModelArk, the vendor's enterprise surface, and not the Dreamina app most readers will actually open. The two surfaces publish different answers on price, resolution ceiling, commercial use, output rights and even whether the product is sold at all in a given country, and only the enterprise one publishes dated documents anybody can re-read. That is honest and it is not the thing most readers will use, so it is stated here rather than buried.

With that fixed: 80.1 out of 100 is the fourth root of 91.6, 79.7, 77.5 and 72.9 multiplied together, and two of those four are not the dimensions they occupy. The card fell when it filled. It read 81.7 until 19 August 2026 on two cells; the two that arrived came in at 79.7 and 77.5, both below the running figure, and a geometric mean is pulled toward everything added to it. Nothing about Seedance changed that day. What changed is that this site stopped leaving rows empty, and on this product the readings that closed them were middling rather than flattering. The arena figure remains the strongest thing on the card, 18,647 blind votes with a published confidence interval.

One movement note matters more than it looks, and it now runs in both directions. On 30 July 2026 this entry read 1225 and sat second; on 12 August 2026 it reads 1,223 and sits third, behind Gemini Omni Flash at 1242 and MiniMax H3 at 1238. Two points of that is the model's own figure drifting and the rest is somebody else's model arriving, which is what makes a published placing decay on activity that has nothing to do with the product. There is a second trap on the same board: Artificial Analysis runs two text-to-video tracks, and this same entry reads 1,223 with audio and 1,265 without. An ELO quoted without its track is ambiguous, and the figure above is the with-audio one throughout. The Artificial Analysis leaderboard is public, so check it on the day you read this. What a row is permitted to publish when its own evidence type does not exist is set out in the scoring model.

Seedance 2.0: What Now Fills the Usability and Price Rows, and What Neither One Measures

Two rows changed their name on 19 August 2026, and the reasons they could not keep it are unrelated to each other. Usability could not be scored because no review sample can honestly be attributed to this model. G2, Capterra, Trustpilot, Product Hunt and Reddit all refuse this desk, a deliberately fake slug draws the identical refusal, so silence here is not a measured zero, and every listing carrying this name that does load belongs to a wrapper rather than to the vendor. Value could not be scored for a different reason entirely: BytePlus does publish prices, per second, stamped by the vendor, and this site's value scale is calibrated in dollars per month. Putting $0.15 a second in the same column as a $10 monthly plan produces a figure that means something different in each row.

So each row publishes a different question with its own name on it, and the distinction is the point rather than a technicality. Access terms reads what the vendor commits to in writing, not what anybody experienced. A vendor with excellent documentation and a miserable product scores well on it. Purchase terms reads what a buyer meets at the checkout, not whether the price is fair. It cannot tell you Seedance is expensive or cheap; that judgement needs the monthly ladder this product cannot be put on.

Two readings behind the purchase-terms figure inverted on contact with the bytes, and both cost the card. The first is the entry commitment. Billing after activation is pay-as-you-go by token usage, which reads like the friendliest possible answer (and it is the billing method rather than the way in. The API document gates the model before any of that: enabling Seedance 2.0 requires a BytePlus account balance over USD 30 or a purchased resource pack with quota available, and the rules document describes a pack as a prepaid resource that offsets token consumption. Two of the packs on sale carry "A minimum of 7 resource packs is required." Collapsing the two would have produced the flattering value, so the reading is a prepaid balance requirement at 62.0. The second is the price publication, which earns the top rung on a fact anybody can check rather than on an impression: three of the vendor's four per-video prices are not the per-second rate multiplied by five. 0.15 by 5 is 0.75 against a published 0.76, 0.37 by 5 is 1.85 against 1.87, 0.78 by 5 is 3.90 against 3.89. The vendor computed the whole-output price and rounded the per-second rate from it, so a reader multiplying would have produced three wrong numbers), which is the difference between a rate card and a rate card with worked examples. The weakest of the four is the resource pack's expiry: "The validity period is 90 days. After it expires, the remaining tokens in the resource pack will be automatically invalidated." Whether the general account balance expires is not stated anywhere captured, and that silence is recorded rather than generalised.

The one thing genuinely left undone is the indemnity, and it is the sharpest loose end on the card. A vendor-side IP indemnity exists in BytePlus's Specific Terms §6.1, with ten enumerated carve-outs, and §1.1 scopes those terms to models "marked as provided by BytePlus on ModelArk", a marking that lives on a console model card behind an account. So whether §6.1 reaches Seedance is not settled by the documents we hold, and a likely answer is not a basis for a published cell. It is also the most heavily weighted of the four positions the commercial-terms dimension asks about, so the 72.9 is an average over what could be read rather than over everything that matters.

Seedance 2.0 Pricing: One Vendor, Two Surfaces, Two Different Answers

Start with the correction, because an earlier version of this page got it wrong in a way worth naming. Dreamina's Seedance 2.0 page publishes no price at all: on the 12 August 2026 capture it sells its plans in credits rather than currency ("Start with free credits, then choose the production capacity and benefits that match your creative needs") and answers the free question the same way in its FAQ: "You can try Dreamina Seedance 2.0 using free daily credits, allowing you to explore its AI video generation capabilities at no cost." That is the whole of the consumer surface's public pricing position. A second correction sits beside the first: this page attributed currently offers free trial credits for eligible users to a 30 July 2026 read of that page, in quotation marks, and neither free trial credits nor eligible users appears in any capture on file. No 30 July capture was kept, so whether the wording changed in the thirteen days between the two reads cannot be shown either way; what can be shown is that those words are not the vendor's today. It does not follow that the vendor publishes nothing. BytePlus ModelArk, ByteDance's enterprise cloud, prints a full rate table stamped by its own page as "Last updated: August 11, 2026", and on a 16:9 five-second clip with no video input it reads $0.07 a second at 480p, $0.15 at 720p, $0.37 at 1080p and $0.78 at 4K, or $0.35, $0.76, $1.87 and $3.89 per video (retrieved 12 August 2026). Those are the vendor's printed figures rather than anyone's arithmetic. Three details from the same page are worth as much as the rates: failed generations are not billed, billing is pay-as-you-go with no monthly or annual view to read the wrong one of, and payment is card-only in US dollars. One qualifier belongs on the middle of those and was added on 19 August 2026, because pay-as-you-go describes what happens after the door opens rather than the door. Enabling the model needs a BytePlus account balance over USD 30 or a purchased resource pack, and a resource pack is prepaid, with two of them sold in minimum lots of seven. What remains genuinely unpublished is the consumer price, and the third-party per-second and per-month figures circulating for Dreamina still come from API resellers, GPU hosts and subscription blogs, most of them undated.

Is Seedance 2.0 Free? What the Trial Credits Actually Buy

Free access to Seedance 2.0 exists in a narrow sense: the vendor's own FAQ says an account can try the model on free daily credits at no cost, and no number is attached to them anywhere on that page (12 August 2026). Daily allowance figures circulate in third-party write-ups, none of which names a vendor source, so no figure is reproduced here. What the credits buy is therefore an evaluation allowance with no published size, and this page does not assign one either. Access is the tighter constraint anyway, since the Creative Partner Program remains invite-only with no public waitlist.

Seedance 2.0 Cost Per Clip: the One Dated Third-Party Figure

The only granular, dated cost breakdown I could find in the public record was published by a competitor. WaveSpeed, a hosting platform that sells alternatives, worked the credit maths out to "roughly $2.50 per 15-second video on fast generation" on top of a $67-a-month advanced tier, with 1080p and 2K output multiplying the burn (6 April 2026). The same review reports that "users report waiting 60-120 seconds for a 5-second clip at standard quality", so a Seedance 2.0 clip is priced in credits and in waiting together. A rival's accounting is not neutral and I am labelling it as such; it remains the most specific dated figure anyone has published for the consumer route. The enterprise route no longer needs a third party at all: BytePlus lists Dreamina Seedance 2.0, Fast and Mini as billable models with the per-second and per-video rates quoted above, so an API buyer can price a clip off the vendor's own page.

Seedance 2.0 Limitations: the Face and IP Filters Commercial Users Now Hit

Two contradictions sit at the centre of this model's record, both dated rather than rhetorical. The first is physics. Launch-week coverage put Seedance 2.0 through fight scenes, fabric and vehicle collisions and concluded that "the model achieves industry-leading levels of generation usability in complex interaction and motion scenes" (Forbes, 12 February 2026). Two months on the practitioner read was narrower: "characters move unnaturally during complex interactions" (WaveSpeed, 6 April 2026), while a Hacker News commenter watching the same clips wrote, "I'm not sure the AI stuff holds together for much more than 15 seconds at the moment" (Hacker News, 15 February 2026). Those accounts are not in factual conflict. One describes a curated demo reel and the other an open-ended workload, and the gap between them is what a buyer actually purchases.

The second contradiction is the filters, and it is the one that changes decisions. The cease-and-desist letters landed within three days of launch, ByteDance tightened face detection and IP matching, and the tightening does not cleanly separate impersonation from legitimate work. A user's own face, licensed characters and stock photography get caught in the same net, which is how a model marketed on realistic, character-driven video ended up described as "dead on arrival" for that exact use (WaveSpeed, 6 April 2026). If character work is the reason you are here, this is the paragraph that decides it.

A third gap is about the reception rather than the clips, and it is narrower than this page once claimed. Elon Musk's one-line reaction on X, "It's happening fast", was reported the day before the public launch (Yahoo Tech, 11 February 2026), and mainstream coverage ran hot from there. On Hacker News the loudest thread was a question about the silence (why nobody there was discussing the model) and the answers were about AI coverage in general, not about Seedance. One commenter said he had stopped clicking on AI stories because they are "mostly hype, slop, fomo, and flat-out lies" (fraaancis, Hacker News, 19 February 2026); another put the quiet down to a crowd "mostly uninterested in AI use" (muzani, same thread, 19 February 2026). This paragraph used to claim more than that. It presented the first of those comments as practitioners staying cold on the same clips, and the commenter was talking about the news cycle rather than about anything Seedance produced. What survives is thinner and still worth knowing: heavy press coverage and a near-empty technical thread six days later, with the dated practitioner readings of the output itself sitting further up this section rather than here.

The community record behind this page is small and I would rather say so than pad it: six themes and fourteen dated quotes from press coverage, Hacker News, one agency and two platform reviews. Nothing came from a Trustpilot page, because none of the Trustpilot pages found reviews the official product, and a widely repeated Reddit thread was discarded because it could not be verified.

Can You Use Seedance 2.0 Commercially? What the ModelArk Terms Answer

This was the least documented thing about Seedance 2.0 until 12 August 2026, and part of it is now answered in the vendor's own words. On output ownership BytePlus is explicit: ModelArk Terms §2.2.10 states that "BytePlus does not assert any ownership of IP Rights in any output generated by the use of any Model Service", and the Specific Terms §3.2 repeat that "BytePlus does not claim ownership of the Output". Read the verbs, because they decide what you can promise a client. The vendor disclaims ownership and permits use; it does not vest ownership in you or assign anything to you, which is why the card scores this at the rights-granted band rather than the ownership band. Money runs the other way: Payment Terms §4.1 makes all purchases "final and non-refundable" unless the law says otherwise, and the ninety-day billing-error path in §4.2 is expressly at the vendor's "sole and absolute discretion", which grants no entitlement. Two things are still open. The free tier states no commercial rule of its own, and the indemnity question in §6.1 is unresolved for the reason set out with the score card. What is documented against all of that is the record of major rights-holders objecting formally within days of launch, answered with stricter filters rather than published assurances, so ByteDance's filters manage ByteDance's risk and the residual exposure sits with whoever delivers the file. Confirm the current terms on the vendor's own page before anything generated here reaches a paying client; these were read on 12 August 2026 and contracts move.

One question lands before the licence question does: what the exported file itself carries. Whether Seedance 2.0 output arrives with a visible watermark, or with a C2PA-style provenance tag written into the file's metadata, and whether either answer differs between free trial credits and paid access, is stated on none of the three official surfaces, Dreamina's product page, the BytePlus product page and the Seed research page, all checked on 30 July 2026, and no dated first-hand account on file settles it. Unknown is what this page records, on the same terms as the unresolved indemnity scope above. If a deliverable has to be unmarked, generate a trial clip and inspect the exported file yourself before you promise a client anything; silence in the vendor's documentation is not a guarantee either way.

Is Seedance 2.0 Safe? the Official Site vs. Unaffiliated "Seedance AI" Clones

Safety here splits into two questions, and the search-side one is more urgent. Looking up the name returns a field of lookalike domains whose trust signals describe other businesses entirely. The Trustpilot page a reader lands on for this name belongs to seedance.ai, an unaffiliated wrapper, so its stars and its complaints say nothing about ByteDance's model in either direction, and the same goes for seedance2.com and the rest of the cluster named earlier. Before entering a card number, confirm the address bar reads dreamina.capcut.com, byteplus.com or seed.bytedance.com.

On the model side one half is now answered and one half is still a labelled gap. The consumer surface carries no review record large enough to score, so nothing on this card describes how BytePlus or Dreamina treats a customer after the money changes hands; what the card does carry is what their documents promise, which is a different question. Regional availability has moved from rumour to text on one point: the reported US exclusion was unverified on 30 July 2026 and is now in the vendor's own words, Specific Terms §1.3, "The BytePlus Model Services are not available in the United States" (read 12 August 2026). That governs the enterprise surface this card scores. The consumer rollout is a separate matter, it began in China, and the launch-country list circulating for it still has no primary source behind it, so it is still not reproduced here.

Seedance 2.0 Alternatives: Which Cells Veo 3 and Kling AI Actually Fill

All three cards now carry four filled rows, so the useful exercise has changed shape: read which of those rows is answering its own axis rather than counting how many are populated. Google's Veo 3 composites 79.0, from an arena ELO read off a board archived on 17 July 2026, a $4.99 entry documented on Google's own plan page, two clause positions in Google's paperwork, and an access-terms reading in the usability position (the same renamed criterion this card carries. Kuaishou's Kling AI is the only one of the three whose four rows all answer their own questions, and it publishes 61.9), a total none of its other rows would predict, because its usability cell is a consumer record of 1.3 out of 5 across 355 Trustpilot reviews. Since 15 August 2026 that record is multiplied in with everything else rather than sitting outside the card and capping the total the way it used to. Seedance 2.0 carries two renamed rows, the most of the three, and no consumer record at all. Both sibling ELOs come off the same 17 July 2026 archive of that board, four weeks before the Seedance figure above, so the table is not a same-day snapshot.

Three composites and three different mixtures. Kling AI's 61.9 used to be produced a different way. A retired rule held any card with a poor, large-sample consumer record at a fixed ceiling, and Kling published at exactly that ceiling (a mechanism's fingerprint rather than a measurement, because it discarded every other cell to get there. Since 15 August 2026 that record costs Kling through ordinary arithmetic instead, on a row that publishes its own n so a reader can weigh it. Read the middle column before the right-hand one. Kling's number is the only one of the three that means, end to end, what the model's four dimension names say it means; the other two include readings of vendor documents standing where a customer record or a monthly price would go. That is not a claim that either is inflated), Seedance's two renamed rows cost it 1.6 points and Veo 3's cost it 1.3, but a total assembled from a different mixture is a different kind of number, and the middle column is there so you can see which you are holding. Every video generator scored here sits on one table, in an order the build computes from those composites rather than one I typed. A direct head-to-head against either model needs both boards and both price pages current on the same day, and it belongs on its own page.

Seedance 2.0 Verdict: Who Should Try It and Who Should Wait

Dated 12 August 2026: on the one independent measurement available, Seedance 2.0 places third of the entries on a public blind-vote board with audio, on 18,647 votes, and most of the rest of what a buyer needs is either split across two vendor surfaces that disagree or disputed by the people using it. Try it if output quality is the question, you can get an invite, and nothing is riding on the result, because the trial credits cost only wait time. Budget it from the BytePlus rate table if you are building on the API, and expect to budget nothing from Dreamina, which still publishes no consumer price. Walk away outright if your work depends on recognizable faces, licensed characters or a licence you can show a client, because that is the workflow the post-cease-and-desist filters now catch, and it is why the earliest practitioners reached for the phrase "dead on arrival". The 80.1 is not a verdict. It is one output measurement, one contract reading and two readings of what BytePlus publishes about buying and keeping access, averaged, on a product where nobody can currently check anything about the customer experience.

Written and scored by Minel Gunesoglu, founder of Vouch · find me on LinkedIn · what this audit does and does not do. Seedance 2.0 arena and contract data collected 12 August 2026, community evidence 30 July 2026; every rank, price and score above carries its retrieval date where it appears. Disclosure: no affiliate relationship exists with ByteDance, Dreamina or BytePlus, so the vendor links here are plain links; elsewhere on this site affiliate links are marked, and a paid submission buys review speed only, never a listing and never a score.