Contents
WAN 3.0 is coming to Picsart. The newest model in Alibaba’s WAN AI video family generates a single unbroken clip of up to half a minute, and it accepts a document or a web page as the source for that clip. Both of those are new. Previous versions stopped at 15 seconds, and took text, images, audio and video as reference and nothing else.
That second change is the one that will alter how a project starts. A deck, a PDF or a product page goes into the model whole. There is no longer a step where all of it has to be boiled down into a paragraph of prompt first.
The rest of the specification: 1080p output, an adaptive aspect ratio, start and end frame control, and rendering at roughly one to two seconds per second of finished video.
Longer takes, without the stitching
A WAN 3.0 clip runs up to 30 seconds, and it runs as one shot. The second half of that sentence matters more than the first. Stitching three clips together leaves joins. At every join the light can shift, the motion can reset and the pacing can stall. One take has no joins.
Smart duration control handles the length. Write the action, and the model works out the tempo implied by it, then proposes a runtime that fits. A slow product reveal wants more room than a fast, cut-driven sequence. The model settles that from the description instead of asking for a number up front.
Finished clips can also be extended. Keep the take that worked and push it further, rather than running the whole prompt again for the sake of a few extra seconds on the end.
The work you already have becomes the brief
Documents go into WAN 3.0 directly, in .doc, .pdf, .ppt and .xls formats. The model reads the file and uses the content as reference for the video. The deck that went to a client, the one-pager, the spec sheet that lists every feature: any of them goes in whole.
Web page URLs work the same way. Give it the address of an article, a research paper, a product page or a marketing site, and whatever sits on that page becomes the reference. A launch page turns into a launch video with no rewrite in between.
Long instructions hold up better too. A brief with four or five separate demands in it reaches the final frame intact, instead of losing its last two clauses somewhere along the way. Writing more pays off here, not less.
Scenes that stay put, and text you can read
Text on screen renders legibly and accurately. Dense, information-heavy scenes gain the most from that, since they carry the highest number of things to get wrong. Price cards, spec callouts, step labels and product names all hang on it.
Detail lands nearer to real footage than earlier releases managed. Nothing drifts either: characters, objects, scenes, styles and audio are all held at pixel level for the length of the sequence. Half a minute tests that far harder than five seconds does, since a face or a product has six times as long to come apart.
Motion, audio and emotion carry more range as well. Pin the first and last frame, and adaptive ratio shapes everything between them.
Where to find it, and what to use until then
WAN 3.0 lands in AI Playground. One prompt runs against several models there at once, so the difference shows up before anything gets committed to.
WAN 2.7 sits in the same place, and it is the WAN model to work with in the meantime. It runs 15 seconds in fixed blocks of 5, 10 or 15, takes up to 5 reference images, and still holds the 4K advantage over WAN 3.0’s 1080p. Reach for 3.0 when length, input range or readable text decides the job. Reach for 2.7 when resolution does.
Running one brief through both, and through the other AI video models sitting in there, is the quickest way to see which suits a given shot.
Five things to line up before it lands
None of this needs the model to be live yet. All five are worth doing now.
- Pull the documents worth handing over. Decks, one-pagers and spec sheets are all valid input. Find the ones that already say the right thing, so they are ready to go in rather than waiting to be rewritten.
- List the pages worth pointing it at. Product pages, articles and campaign sites all work as reference. A launch page that is already signed off is the fastest route to a launch video.
- Write the briefs longer. Detail survives to the final frame now. Anything trimmed out of a prompt to keep it short is worth putting back in.
- Storyboard for one take. A 30-second idea no longer has to be planned as three clips with hidden joins. Plan the whole thing as a single shot instead.
- Choose the opening and closing frame. Both can be pinned, so decide where a clip starts and where it ends, and let the model fill in the motion between them.
What is new in WAN 3.0
Six things separate WAN 3.0 from the releases before it.
| What changed | What it means |
|---|---|
| Clip length | Half a minute, uncut, where earlier releases stopped at 15 seconds |
| Smart duration | The runtime comes from the described pacing, instead of being picked up front |
| Extend | Any finished clip can be extended into a longer sequence |
| Reference inputs | Documents and web page URLs join text, image, audio and video |
| Text on screen | Words render legibly and accurately, including in dense, information-heavy scenes |
| Reference consistency | References hold at pixel level, so nothing drifts out of shape over a longer clip |
The short version: earlier releases generated video in blocks. WAN 3.0 generates a scene, and lets material that already exists decide what goes in it. For how the newer model measures up against the one in Picsart today, spec by spec, read WAN 3.0 vs WAN 2.7.
Get answers to common questions
WAN 3.0 is the newest model in Alibaba’s WAN AI video family. It makes a single unbroken clip of up to half a minute at 1080p, shapes it to an adaptive aspect ratio, and lets the opening and closing frame both be fixed. Six kinds of reference feed it, including documents and web addresses.
Get ready for 30 seconds in one take
WAN 3.0 brings the length, the input range and the readable on-screen text, and it lands in AI Playground. Start gathering the documents and product pages worth feeding it, and run the same briefs through AI Playground today to see where the current models land.
