Most write-ups about generative video stop at "it looks good". This page is for the person downstream of that — the one who has to drop the file on a timeline that already has a frame rate, a raster and a delivery spec attached to it.
Three numbers decide whether that is a five-minute job or an afternoon.
Wan 3.0 outputs 30 fps. Not 24, not 25, not variable.
If your timeline is 24 fps — and if you are cutting anything that will be graded or intercut with cinema camera footage, it is — every generated clip has to be conformed. Your options are the usual three, and they are all compromises:
None of this is a defect. It is a spec, and specs are only a problem when you find them after you have cut the sequence. Set the sequence to 30 and conform at the end, or start at 24 and accept the conform on every clip.
At 25 fps for European broadcast the arithmetic is friendlier — a 30 to 25 pulldown is less visible than 30 to 24 — but it is still a conform.
The ratios available are adaptive, 16:9, 4:3, 1:1, 3:4 and 9:16.
The one that is not there is 21:9, and it is worth being explicit because a number of pages describing this model list it. There is no cinema ratio. If you need 2.39:1 you are cropping 16:9, which costs you vertical resolution you may not have to spare — and you should decide that before you generate rather than after, because framing for a crop changes what you ask for.
The vertical ratios are real, though, and that matters more than it sounds: 9:16 comes out of the same model at the same length ceiling, so a vertical cut is a re-generation rather than a re-frame. Whether that is cheaper than cropping depends on your rate.
480P, 720P, 1080P. 1080P is the ceiling. There is no 4K tier, and anything sold above 1080P for this model is an upscaler running after the fact — a different process with a different failure mode, and not something the model produced.
Two working notes:
The 480P tier is new to this generation and it is the one that changes daily practice, because a draft costs a quarter of a final. Iterate at 480P, spend once at the end.
Omit the resolution field and you get 1080P, which is four times the 480P rate. On a per-second product, the default is a financial parameter, and this is the single most common way an estimate comes back four times low.
Integers only, and thirty is the ceiling for a single generation. The important part for an edit is that thirty seconds arrives as one continuous take rather than several clips joined — so there are no internal cut points to reuse, and no way to replace a section without regenerating the whole thing.
If a video goes in as input, its duration counts against the same thirty seconds and is billed at your output rate on top of the output seconds. Trim the source to the last few seconds before sending it.
Speech, effects and music are generated alongside the picture in a single pass. Two consequences for post:
Plan accordingly: if the spot needs a licensed music bed and a mixed VO, generate with audio off in mind only for reference, and lay your own audio underneath. If the point of using this model is the same-pass sound, then the mix you get is the mix.
| Spec | Value | Downstream consequence |
|---|---|---|
| Frame rate | 30 fps | Conform for 24/25 fps timelines |
| Widest ratio | 16:9 | 2.39:1 is a crop, not an output |
| Resolution | 480P / 720P / 1080P | No 4K; default is 1080P at 4× the 480P rate |
| Length | 2–30 s, integers | One continuous pass, no internal cuts |
| Audio | Same pass, mixed | No stems, no discount for switching it off |
Each of these rows is a published constraint with a source and a check date on the Wan 3.0 specification page. If you want to put a real file on a real timeline before committing to any of it, a free Wan 3.0 clip on the real model needs an email and no card.
wan-3.run is an independent third-party interface built on Wan 3.0. It is not affiliated with, endorsed by, or sponsored by Alibaba Group or Alibaba Cloud.