Wan 2.7 · Text to Image

Here's the honest pitch for this node: it's a sketchbook. Four images per prompt, 1K or 2K, 4.5 credits ($0.045) each, with an optional reasoning pass in front. That's a few cents a candidate for a model whose fine-detail work is its strong suit - and the community's one real evaluation of Wan 2.7 singled out exactly that: best small-detail textures of the models it was compared against on close-ups.

It is, of course, closed source. Wan stopped being downloadable at 2.2, and the 2.7 release in April 2026 was met with a shrug so loud it became its own thread. Whether that matters depends on what you want: if you want to own the model, this isn't your node. If you want a handful of high-detail candidates without babysitting a sampler, it's a reasonable trade.

How it works

The node is generated from the model's input schema, so the widgets are the endpoint's real fields. Queue it and the pack quotes the run before holding credits, submits with an idempotency key, streams status onto the node, and downloads the results as one batched IMAGE. Four outputs means a batch of four - Save Image will write four files, and you can index the batch downstream if you only want the winner.

enable_thinking is the phase everyone asks about. It runs a deep-thinking pass before generating, which the tooltip says makes the output better and the wait longer, and it's on by default. Turn it off when you're iterating on composition and you'd rather see six rough ideas than two considered ones.

The inputs that matter

prompt - subject, composition, style, lighting. There's no negative prompt field on this model, so anything you want excluded goes into the prompt itself as a positive statement ("plain backdrop, no text").

num_outputs - 1 to 4, billed per image returned. This is the field that decides whether this node is a sketchbook or a slot machine.

resolution - 1K or 2K. aspect_ratio - 1:1 default, plus 4:3, 3:4, 16:9, 9:16, 3:2, 2:3.

color_palette - a JSON string of 3 to 10 colours with optional percentage shares. It looks like a gimmick and it isn't; it's the most direct control this pack gives you over the look of a render, and it's far more effective than stacking adjectives.

seed sits at the bottom doing the usual job: fixed by default, so queueing the same graph twice returns the same images rather than charging you again.

Outputs: images and task id.

Installing it

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

Or search PopNow in ComfyUI Manager and install from there. Restart, add your key from popnow.ai/api/console/keys in Settings > PopNow (or set POPNOW_API_KEY on a server), restart again. The pack is a thin HTTP client - no checkpoints, no VAE, no Python dependencies beyond what ComfyUI already ships, and it runs fine on a machine with no GPU.

Delete custom_nodes/popnow_api.py if you still have the old single-file node.

Where people get burned

color_palette takes JSON as a string; malformed JSON fails the run rather than degrading gracefully. Same story for anything you template in from upstream.

The thinking pass is a budget decision, not a quality switch you leave on out of caution. It costs you wall-clock time and it doesn't make a badly-written prompt good. Write the prompt properly first, then decide.

And the pack-wide gotcha: interrupting a run doesn't cancel it. The task keeps generating server-side and settles if it succeeds. If you're auditioning prompts, Settings > PopNow > Max credits per run is the actual safety rail - the node refuses anything quoted above it before credits are held.

Code reference, since these come up: 40201 credits, 40004 unsupported settings (field named), 400 rejected value, 40310 unverified email, content_rejected declined and refunded.