Z-Image · Text to Image

From 1.5 credits an image. At 1 credit = $0.001 that's $0.0015, and yes, I had to check the decimal point too. A thousand generations costs about a dollar fifty. If you want to know why the Z-Image line is worth caring about - 6B parameters, Apache 2.0, better skin than the Flux defaults of its era - this is the least expensive way to find out that exists.

The plain version of the pitch

Z-Image Turbo is Alibaba Tongyi-MAI's distilled image model: flow-matching, 8 NFEs, CFG 1, trained on bilingual data. The reaction when it landed in late 2025 wasn't polite interest; it was people posting 3060 timings. It slotted into the local ecosystem because it's small, permissively licensed and genuinely fast, and it held up next to much heavier models on portraits.

This node is that model with none of the setup. No checkpoint, no VAE, no text encoder, no sampler. You type, you queue, you get a JPEG-quality PNG back on a node that takes up no VRAM. It's the node I'd hand someone who owns a laptop and wants to know whether all this is for them.

What's on the node

prompt - subject, composition, style, lighting. Sentences work better than tag soup on this family.

aspect_ratio is the short list: 1:1, 4:3, 3:4, 16:9, 9:16. No 21:9 here, which is a real difference from the LoRA variants. If you need an ultrawide, generate 16:9 and crop, or use a LoRA node.

enable_prompt_expansion defaults off, and there's a money reason to leave it there: the tooltip says requests with it on are billed at the prompt-expansion rate. You're paying for someone to rewrite your prompt before rendering. Occasionally worth it on a three-word idea. Mostly not worth it when you already know how to write.

output_format is only offered as png on this model - that's what the catalog advertises - so don't plan on a webp pipeline. seed defaults to fixed, which means the same graph queued twice doesn't quietly cost you a second image.

There's no loras input on this one. That's the main reason it exists next to the Turbo LoRA node: this one is the plain model, cheapest possible. If you want your own weights, use the LoRA variant and pay the extra three-tenths of a credit.

Outputs are images (an IMAGE batch, straight into Save Image, an upscaler, or a compositing chain) and task id for the API console.

What happens on queue

The node's fields are generated from alibaba/z-image-turbo/text-to-image's published schema, so nothing is invented. It builds the request, skips anything left at "model default", asks PopNow for a quote before holding credits, submits with an idempotency key so a retried request can't create a second billed task, polls the job, then downloads and decodes the image. The node prints the estimate up front and the final charge when it's done, which is more transparency than most metered tools manage.

Worth knowing: Z-Image was trained bilingual, and Chinese prompts get a measurable adherence boost. If a scene keeps slipping, that's a trick locals have been using since release, and the API doesn't care what language you type.

Install

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

Or Manager → PopNow → install. Restart ComfyUI, create a key at popnow.ai/api/console/keys, paste it in Settings > PopNow or set POPNOW_API_KEY, restart once more so the node list is rebuilt from your account. Nothing to download, no GPU required, no Python dependencies beyond ComfyUI.

Where people get burned

Prompt expansion is the sneaky one: it's advanced-flagged, easy to flip on, and it changes both your prompt and the rate. If your seed suddenly "doesn't reproduce", check whether a rewrite is happening before the render.

Second, this node gets you images, not control. No LoRA, no mask, no step count. If you're coming from a local Z-Image workflow with a sampler and a scheduler you liked, you'll miss them; the trade is that you also can't get them wrong.

Third, cheap doesn't mean free of policy. The model still refuses what it refuses, a content_rejected run just doesn't cost you anything, and prompts leave your machine - that's the deal with every API node in this pack.