Wan 3.0 vs Seedance 2.5: the 4K everyone quotes did not ship

These two get compared constantly, and for once the community got there before the press — when Wan 3.0 was announced, Seedance was the model people immediately named as the thing it had to beat.

Two facts reshape that comparison and both are checkable in a minute. Seedance 2.5 was announced with native 4K and did not ship it — ByteDance's own channel documents 480p and 720p, and its predecessor Seedance 2.0 actually reaches higher. And ByteDance has not published a per-second rate for 2.5, which is a planning problem rather than a quality claim, but it is a real one.

Meanwhile Seedance 2.5 genuinely beats Wan 3.0 on one axis by a distance, and that axis is references.

The parameter tables

ByteDance's figures as documented on the Volcengine channel and as they appear in the shipped Seedance 2.5 request schema; Alibaba's from the Wan 3.0 API reference. Both read 2026-08-25.

The 4K that did not arrive

Seedance 2.5's announcement promised native 4K in a single pass, and that promise is still being repeated in comparison tables written months later. The shipped model documents 480p and 720p. There is no 1080p tier on ByteDance's own channel, let alone anything above it.

That produces the odd situation where Seedance 2.0 outranks Seedance 2.5 on resolution — 2.0 has a 1080p tier and 2.5 does not — so "the newer model" is not automatically the higher-resolution one. Some resellers do surface a 1080p option for 2.5, which is worth verifying against whichever channel you are actually calling before you promise a client anything.

Set against that, Wan 3.0's 1080P is the higher documented ceiling of the two. Neither reaches 4K, and any table telling you either of them does is repeating an announcement rather than a specification.

Where Seedance 2.5 wins

References, decisively. Thirty images, ten video clips and ten audio clips — fifty assets against Wan 3.0's twenty, with thirty seconds of reference video allowed against fifteen. If your work is built on a large cast, a product catalogue or a deep style library, this is not close and Seedance is the answer.

21:9. Seedance offers the cinema ratio; Wan 3.0's widest is 16:9. For letterboxed work that decides it on its own.

Editing and extension as first-class task types, including extending an existing clip beyond its original length. Wan 3.0 edits but does not extend, so anything past thirty seconds is a fresh generation either way — Seedance gives you a continuation path.

Where Wan 3.0 wins

A published price. Alibaba lists $0.05, $0.10 and $0.20 per second at its three tiers. ByteDance has not published an equivalent for 2.5, so every number you find for it comes from a reseller with its own markup. You cannot budget against a rate that does not exist, and for anyone planning a month of output rather than one clip, that is the difference between a forecast and a guess.

The higher documented resolution. 1080P against 720p, per each vendor's own channel.

Documents and web pages as input. A deck, a PDF, a spreadsheet or a public URL becomes source material. Seedance has no equivalent, and neither does anything else in this category. What that actually does.

Audio at no premium, stated in the parameter table rather than inferred.

A 480P tier. Both models have a cheap tier, and Wan 3.0's is the cheapest place to be wrong in this comparison at a published five cents a second.

A lower duration floor. Two seconds against four. Trivial until you are making a loop or a transition, at which point four seconds is twice as much clip as the shot needs.

One place they behave identically, and it catches everyone

Both models bill reference video seconds on top of output seconds. Attach a reference clip and its duration is added to what you pay, at your output rate, on both platforms. Reference images and audio are free on both.

This matters more on Seedance precisely because its reference budget is so much larger: thirty seconds of allowed reference video, all of it billable, against Wan 3.0's fifteen. The bigger reference budget is a real capability and a real exposure at the same time, and it is the one place where "more" is not unambiguously better.

Both also accept a model-chosen duration, and both reserve at the thirty-second ceiling when you use it. Name a length on either. The billing arithmetic in full.

Three claims worth checking

Four briefs

The shape: Seedance 2.5 is built for briefs with a lot of source material. Wan 3.0 is built for briefs that need a finished-resolution deliverable and a price you can plan against. They overlap on length and diverge almost everywhere else, which is why the honest answer depends on what your inputs look like rather than which model is "better".

Questions

Does Seedance 2.5 support 4K?

No. It was announced with native 4K and shipped at 480p and 720p on ByteDance's own channel. Wan 3.0 does not reach 4K either — its ceiling is 1080P, which is still the higher of the two.

Which is higher resolution, Wan 3.0 or Seedance 2.5?

Wan 3.0, on published specifications: 1080P against 720p. Unusually, Seedance 2.0 also has a 1080p tier that 2.5 does not, so the newer Seedance is not the higher-resolution one.

How much does Seedance 2.5 cost?

ByteDance has not published a per-second rate for it. Every figure in circulation comes from a reseller. Alibaba does publish Wan 3.0's — $0.05, $0.10 and $0.20 per second — which is worth weighing if you are budgeting rather than experimenting.

Which takes more reference material?

Seedance 2.5, by a wide margin: 30 images, 10 videos and 10 audio clips against Wan 3.0's 10, 5 and 5. Note that reference video seconds are billed on both, so the larger budget is also a larger bill.

Can either one take a PDF or a deck?

Only Wan 3.0. Document and web-link input has no Seedance equivalent.

Which is longer?

A tie in practice. Both reach 30 seconds in one generation; Wan 3.0 starts at 2 seconds and Seedance at 4.

---

The advantage worth acting on here is a price you can plan against. Wan 3.0's rate card fits on one page, the arithmetic is per second rather than per credit, and the first clip does not touch it.