Wan 2.6 · Text to Video

Text to video is the task that makes the most sense to outsource. There's no input to keep on your own disk, no frame you're precious about - you type a scene and something either useful or useless comes back. On a local Wan 2.2 setup, every one of those experiments costs you minutes of denoising. Here it costs seconds of queue time and a few credits.

This is Wan 2.6's text-to-video, the full tier: 720p or 1080p, 5 to 15 seconds, single or multi-shot, from 13.5 credits ($0.135) per second of output. The model has no public weights - Alibaba has shipped nothing open since Wan 2.2 - so an API node is the only way in.

How it works

The pack generates this node straight from Wan 2.6's input schema; the widgets mirror the endpoint exactly, including the option lists. Queue a prompt and the node quotes the run first, submits the task under an idempotency key, streams progress onto the node, then downloads the finished MP4 as a ComfyUI VIDEO.

One design detail that will save you money: the seed starts on fixed, not randomize. Queue the same untouched graph twice and ComfyUI hands you back the first result instead of quietly billing a second run. Switch control-after-generate to randomize only when you actually want a new roll.

The inputs that matter

prompt is the whole thing. The tooltip asks for subject, action, camera, light, pacing and boundaries - which is a decent checklist if you're staring at a blank box. That last word is doing work: telling the model where the shot ends is how you stop it from wandering off in the final second.

shot_type is the interesting control. single is one continuous take; multi lets it cut, which is why this node can produce something that reads like a sequence instead of a loop. With multi, write the prompt as beats - wide, then close, then the reveal - and let the model place the cuts.

duration_seconds snaps to 5, 10 or 15. resolution is 720p or 1080p. aspect_ratio gives you 16:9, 9:16, 1:1, 4:3 or 3:4, so vertical and square are first-class rather than an afterthought.

negative_prompt is worth the ten seconds it takes. enable_prompt_expansion will rewrite your brief into something fuller before generating - fine for terse prompts, counterproductive when you've already written exactly what you want.

There's also an optional audio_url socket: an AUDIO track the clip follows for lip sync and pacing. Wire a voiceover in and the mouth shapes follow it. That's a sync track, not a soundtrack.

Outputs: video, plus task id for the API console logs.

Installing it

Manager > Custom Nodes Manager > search PopNow > Install > restart. Or the git route:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Then create a key at popnow.ai/api/console/keys and paste it into Settings > PopNow, or set POPNOW_API_KEY. Restart again after the key goes in - the node list is built from your account's catalog at startup. Nothing else downloads: this is a thin HTTP client, not a local inference stack, and it'll run happily on a laptop with no GPU at all.

If you're migrating from the old single-file node on popnow.ai's docs page, delete custom_nodes/popnow_api.py.

Where people get burned

The bill. Fifteen seconds of clip at four times the Flash rate is a real number, and that's before you generate a second take because the first one had six fingers. Set Settings > PopNow > Max credits per run and let the quote gate do the worrying.

Prose-style prompting is the other trap. Text-to-video models don't reward adjectives the way image models do; they reward geometry and time - who moves where, which way the camera goes, what happens at second four.

Error codes: 40201 not enough credits, 40004 unsupported settings combination (it names the field), 400 a rejected value, content_rejected means the model declined and you weren't charged. Missing model in the node search? Restart - the catalog is fetched at startup, and a mid-session key change doesn't rebuild it.