Wan 2.2 Review: Open-Weight Motion, One Path Uncensored

We don't publish our content-ladder runs; the gallery below is the everyday register, taken from the published, safe-for-work examples on the Wan 2.2 model page.

A shape behind gauze. A wind chime turns and a silhouette lying behind a curtain shifts with it — the whole clip is built from one still and a prompt describing cause and effect, nothing more.

A dropper bottle, lit by a moving window. The camera arcs a slow quarter turn while a band of light slides across the glass; the bottle itself never moves, which is the point of the test.

A step forward in a grey studio. From a portrait still, she turns her shoulders and takes two slow steps, satin catching the light — a 720p take that shows how far a single gesture can carry a five-second clip.

Two people wrapped in one coat. Written from nothing at 720p: snow falls on an empty platform while two people share one overcoat, breath clouding, a lamp flickering — one continuous, cinematic beat.

A lighthouse stair, climbed at dusk. A 480p text-to-video draft: one action, one camera position, checking whether a shot idea holds up in motion before spending more on resolution.

A night market after rain. A vertical 720p take with a fixed seed, so the same framing can be compared across prompt edits — steam, bulb strings and a wet street reflecting neon as she walks toward camera.

On image-to-video, the still fixes who, what they're wearing and where they are; the prompt should carry only the motion and, where it matters, a cause for that motion. On text-to-video, name the cast directly and keep the count and colour specific — vague crowd or ensemble language is where we saw the most drift.

1. Cause and effect from a still (image-to-video).

Cast: one woman. The camera stays on the shape behind the gauze. A wind chime turns and the figure lying behind the gauze curtain shifts with it, the curtain's shadow travelling across the body. Silent. Cinematic, heavy film grain.

Run this prompt on Wan 2.2 →

2. A named cast and a specific count (text-to-video).

Cast: one man and one woman. The camera stays on the two of them inside the one coat. Snow falls on an empty platform; two people are wrapped inside one overcoat, breath clouding between them, the platform lamp flickering. Cinematic, heavy film grain, sodium lamp in snow blue.

Run this prompt on Wan 2.2 →

Weak: "A ballerina practises alone in a bright studio." Left this open, our own test came back with three dancers instead of one. Strong: name the cast up front — "Cast: one woman" — and repeat the count inside the sentence describing the action, the way prompt 2 repeats "two of them" and "two people."

Between 16 and 25 September 2026 we ran our five-rung content ladder on both Wan 2.2 endpoints. On image-to-video, L1 Suggestive, L2 Mature themes and L3 Strong rendered clean on every one of our runs; L4 Intense came back softened on all three attempts; the L5 Extreme restraint/impact probe rendered once in four tries and softened the rest; the L5 Extreme gore probe rendered clean on all three runs. On text-to-video, every rung we sent — 19 runs in total, from L1 Suggestive through both L5 Extreme probe types — came back refused.

The page labels reflect that split: image-to-video carries the softened badge, text-to-video the filtered one. We add no content filter of our own on either endpoint; this is Alibaba's own policy behaviour on each task.

Not tested yet: whether a closing frame or a longer, worded prompt changes the L4 Intense softening rate on image-to-video; scenes with more than two people at the harder rungs; anime or illustrated styles on the ladder.

Adults only, in every prompt, still and output. Never depict a real, identifiable person without their consent, and make sure you hold the rights to any photo you upload as an opening frame.

Wan 2.2's reputation outside our own tests is built on it being Alibaba's last open-weight release in the line: later versions (2.5 onward) are API-only. That decision still draws complaints over a year later. One Hacker News commenter, frustrated that Alibaba stopped publishing weights for anything newer, put it directly: "Wan 2.2 did a rug pull, left a huge gap for the community" that had built tools and fine-tunes on top of it.

Our own runs land in the middle of that story: the open weights behind this endpoint still animate a still convincingly on everyday and moderately strong prompts, and the same openness is presumably why the image-to-video path tolerates more than the text path does.

• Bringing your own trained weights? Wan 2.2 LoRA runs the same base with up to three LoRA slots, by the second rather than the flat per-clip price. Our Wan 2.2 LoRA review.
• Want the uncensored version priced by the second? Wan 2.2 Spicy is image-to-video only, $0.019 a second at 480p, and adds an eight-second option and a closing frame.
• Need motion transfer or character animation instead of a caption or a static still? Wan 2.2 Animate drives a character from a reference performance.
• Want newer motion and a longer take? Wan 3.0 adds multi-shot direction and clips up to 30 seconds, priced by the second from $0.045 at 480p.