Wan 2.7 Pro · Text to Image

The one genuinely useful thing anyone has measured about Wan 2.7 is that it renders small detail better than its peers - the only substantive comparison in the community corpus put it ahead of HappyHorse, Veo 3.1 and Seedance on close-up texture: faces, fabric, water. That's a still-image virtue wearing a video model's clothes. So if you're going to spend on the Wan 2.7 family at all, spending it on a 4K text-to-image render is the least silly way to do it.

This is the Pro tier of that: 2K or 4K output, seven aspect ratios, and a color_palette field that lets you hand the model a real colour scheme. 11.25 credits ($0.1125) per image - about a tenth of a cent per credit, so roughly eleven cents per render.

And no, you can't run it locally. Nothing past Wan 2.2 has public weights.

How it works

Like every node in the pack, this one is generated from the model's published input schema, so the fields and their option lists are the endpoint's own. Queue it and the node quotes the job, creates the task under an idempotency key, polls with progress text on the node, and downloads the finished image into ComfyUI as a normal IMAGE. There's nothing to keep in VRAM afterwards.

Note the absence of a negative prompt: this model doesn't take one. The prompt tooltip says it plainly - write what to leave out directly in the prompt. That's a real adjustment if you're used to bolting a 200-word negative onto every graph.

The inputs that matter

prompt - subject, composition, style, lighting, plus any wording you want rendered into the image. 2.7 is decent at legible text, which is more than you can say for most of the field.

resolution - 2K or 4K. Both cost the same per the node's own tooltip, so the only reason not to take 4K is turnaround time.

color_palette - this is the field that makes the node interesting. A JSON string of 3 to 10 colours, each with an optional share of the image in percent; leave the shares out and it splits them evenly. It's a crude form of art direction and it works far better than three paragraphs of "muted teal and rust, cinematic".

aspect_ratio - 1:1 by default, with 4:3, 3:4, 16:9, 9:16, 3:2 and 2:3 available. Pick it deliberately; a 4K square and a 4K 16:9 are very different amounts of pixels.

seed behaves like the rest of the pack: fixed by default, so re-queueing the same graph reuses the result rather than re-billing you.

Outputs: images and task id. One image per run - there's no count widget on this tier.

Installing it

Manager > Custom Nodes Manager > search PopNow > Install > restart ComfyUI. Or:

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

Then get a key at popnow.ai/api/console/keys, paste it into Settings > PopNow (or export POPNOW_API_KEY), and restart so the node list rebuilds from your account's catalog. Nothing else to install - no checkpoints, no VAE, no custom samplers.

If you're on the single-file popnow_api.py from popnow.ai's docs, delete it before you start.

Where people get burned

color_palette is a string field expecting JSON. A trailing comma will fail the run before anything is generated, which is at least a cheap failure. If you're building the palette dynamically, format it upstream and pass the string in.

The other trap is more conceptual: 4K doesn't buy you more prompt understanding. If the composition is wrong, upscaling the request won't fix it - iterate at the cheaper tier or on the cheaper 1K Wan 2.7 node, then take the winning prompt to Pro.

Error codes you'll actually see: 40201 (out of credits), 40004 (unsupported combination, field named), 400 (rejected value), content_rejected (the model declined; refunded). And if a model seems to be missing from the node search, restart ComfyUI - the catalog is fetched at startup.