Wan 2.2 · Text to Video

What it is

Wan 2.2 text-to-video, base model, no adapters, no duration widget. The node's own description calls them "simple five-second clips at 480p or 720p," and that's a promise rather than a warning: this is the least fiddly way to get moving pictures out of the model the open-video scene spent two years standardising on.

It's worth being clear about what Wan 2.2 is, because it shapes everything you'll get out of this node. A 27B mixture-of-experts model, 14B active, with a high-noise expert for motion and composition and a low-noise expert for detail. Apache 2.0. And frozen - Alibaba moved 2.3 and everything after it behind an API, so 2.2 is the last open Wan base that will ever exist. Running it locally means 12GB of VRAM at a squeeze, two-pass sampling, and a clip every few minutes on a good card. This node skips all of that.

How it works

There's no sampling on your machine. The node assembles a request from five fields, PopNow quotes it, runs both expert passes, downloads the MP4, and returns a ComfyUI VIDEO value. Point it at Save Video - Save Image will not accept a VIDEO socket, which is the number one wiring mistake people make with video nodes.

Progress appears as text on the node: the estimate, the queue position, the time taken, and finally the charge. Each run is quoted before anything is held, and failed runs are refunded.

The inputs that matter

prompt - subject, action, camera, light, pacing, and where the shot should end. The tooltip's list is not padding; with a fixed five-second window, "and boundaries for the shot" is doing real work. A clip with a stated ending feels intentional. A clip without one feels cut off.

Wan's prompting taste, from the people who run it locally: prose, not tags, with motion described plainly. "A cyclist cuts across a wet intersection, brakes late, and the camera tracks right as the light changes" gives the model a sequence. Three adjectives and a noun give it a vibe, and vibes don't move.

aspect_ratio is 16:9 or 9:16, nothing else. resolution is 480p or 720p, and 720p doubles the per-run price - so draft at 480p, because motion either works or it doesn't and you'll know immediately.

enable_prompt_expansion is off by default and I'd keep it off here. On a text-to-video endpoint the rewriter can help a thin brief, but it also rewrites the camera and the ending, which are the two things you were being specific about.

seed defaults to fixed. Re-queueing an identical graph gives you the identical clip; switch control-after-generate to randomize for a new take, and remember each randomised run bills.

Outputs: video, task id.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui   # or Manager > search "PopNow"

Create a key at popnow.ai/api/console/keys, restart, paste it into Settings > PopNow (or set POPNOW_API_KEY for a headless install), then restart once more so the node list is fetched from your account's catalog. No model files, no CUDA, no pip installs - the pack ships with an empty dependency list on purpose.

Where it goes wrong

Faces and hands are Wan 2.2's known weak spots, and a five-second clip with a lot of action in it gives them maximum opportunity to wobble. Fewer subjects, less movement, one camera move. If you need a specific face to survive the shot, start from a still instead - that's what the image-to-video version of this node is for.

If you want a longer clip, this node structurally can't give you one: five seconds is the endpoint, because 81 frames is the model. Use the Video Extend node to continue from the last frame.

The pack errors are consistent across nodes: PopNow 401 is a bad key, 40201 means an empty wallet, 40004 means the model can't serve that combination, content_rejected means it declined and you were refunded.