Every specification card for this model describes what comes out: three resolution tiers, six aspect ratios, 30 fps, two to thirty seconds. Those are the numbers people quote, and they are the numbers that almost never cause a failure.
The failures are on the way in. Here is that side of the sheet.
| Formats | JPEG / JPG / PNG without an alpha channel / BMP / WEBP |
| Each side | 240 – 8000 px |
| Aspect ratio | ≤ 8:1 |
| File size | ≤ 20 MB |
Three of those bite in practice.
Transparency is not supported. A PNG with an alpha channel is a normal, correct PNG everywhere else in your pipeline. Logo assets, product cut-outs and anything that came out of a design tool are exactly the files most likely to carry one, and they are also exactly the files people reach for as a reference image.
There is a floor, not just a ceiling. 240 px on the short side rules out thumbnails and most avatars. If your app lets a user pick an image they already uploaded, the version your CDN hands back may be a resized derivative that is under the floor while the original was fine.
8:1 is about the input, and it is not the output ratio list. A tall product shot or a wide banner crop can exceed 8:1 without looking unusual to a person. This is a separate rule from the six aspect ratios the model can produce, and satisfying one tells you nothing about the other.
| Formats | mp4 / mov |
| Each clip | 1 – 15 s |
| Each side | 240 – 4096 px |
| Aspect ratio | ≤ 8:1 |
| File size | ≤ 100 MB per clip |
| Count | up to 5 clips, 15 seconds in total |
Note the two numbers that look the same and are not. A single clip may be up to 15 seconds, and five clips are allowed — but the total across all of them is also 15 seconds. Five fifteen-second clips is not a legal request; five three-second clips is. "Up to 5" and "up to 15 s" are frequently quoted next to each other as though they multiply.
Also note the ceiling is 4096 px, lower than the 8000 px allowed for stills. A 6K frame grab is a legal image and an illegal video.
| Formats | wav / mp3 |
| Each clip | 1 – 15 s |
| File size | ≤ 15 MB |
| Count | up to 5 clips, 15 seconds in total |
Same shape as video: five slots, one shared fifteen-second budget.
| Formats | docx doc xlsx xls pptx ppt pdf txt key pages numbers md |
| Size | ≤ 100 MB |
| Pages | ≤ 50 — validated for pdf, docx, doc, pptx, ppt, key, pages only |
| Count | 1 file, or 1 link, never both |
That parenthetical is worth reading twice. Twelve formats are accepted; the page limit
is checked on seven. xlsx, xls, txt, numbers and md are checked for size and
nothing else — which means an enormous one of those is not rejected, it is accepted.
Links must be publicly accessible pages that do not require login. An internal wiki URL is not an error you will be warned about at request time.
Three transports, all equivalent as far as the constraints above are concerned:
https://… a public HTTP(S) URL
oss://… an OSS temporary URL
data:{MIME};base64,{data} inline base64
Base64 does not exempt anything from the size caps. It inflates the payload by roughly a third on the wire, so a 20 MB image is closer to 27 MB of request body.
Every rule above produces a rejection after the user has committed — chosen a file, written a prompt, pressed the button. And because only two jobs run at a time on this model, a rejection is not just an error message; it is a queue slot and a wait spent on nothing.
All of it is checkable client-side before a single byte is uploaded. Image dimensions and alpha channel come out of the file header, clip duration and pixel dimensions out of the container metadata, and the fifteen-second totals are addition. The most expensive version of this page is the one where you learn the rules one production incident at a time.
Formats, dimensions, durations, file-size caps and the transport list are from Alibaba
Cloud's Model Studio API reference for wan3.0-video, read 2026-08-27. The parenthetical
limiting page validation to seven of the twelve document formats is quoted from that
same reference.
The output side of the sheet — the six aspect ratios, the three resolution tiers, 30 fps, and the two shapes people expect and do not get — is the frame card on this site, and the shape a request takes once these files are attached is the translation card.
If checking the rules by hand is not how you want to spend an afternoon, the reference mode, which takes every input type on this page will tell you what it accepts faster than the table will. Disclosure: wan-3.run is my site.