Z-Image Turbo LoRA · Text to Image

If you've trained a Z-Image Turbo LoRA - or downloaded one off Civitai - this is the shortest path from that file to an image. Point a PopNow LoRA node at it, connect the socket, write your trigger phrase first, pay 1.8 credits. No 6B checkpoint to download, no text encoder, no adapter, no sampler sweep. For a model whose whole selling point was being cheap and quick, that's the right shape of tool.

What Turbo is, in one paragraph

Z-Image Turbo is Alibaba Tongyi-MAI's distilled 6B image model, flow-matching, 8 NFEs, Apache 2.0. It landed in late November 2025 and immediately ate Flux 2 Dev's launch week - the reaction was less "nice model" and more "wait, that was half a minute on a 3060". Its known weakness is sameness: seeds move less than you'd like, and it tends toward conventionally pretty. Its known superpower is that it trains fast, which is why there's an enormous pile of Turbo LoRAs for it, quality varying wildly.

That variance is worth remembering when a LoRA produces mush. The KB's read on the early Turbo LoRA flood was that most bad ones came from low-quality or AI-generated training sets and bad hyperparameters, not from the model being un-LoRA-able. One trainer's "I've made about a dozen Z-image LoRAs... one of the easiest models to train LoRAs for" sits right next to a complaint thread asking why they're all bad. Both are true.

What you set

prompt - subject, composition, lighting, and if a LoRA has a trigger word, first. That ordering advice is the author's, and it's not superstition; early tokens carry more weight.

aspect_ratio covers the usual nine, from 1:1 through 21:9 and 9:21. output_format is jpeg, png or webp. seed starts fixed on purpose, so queueing the same graph twice reuses the first result instead of billing a second - flip control-after-generate to randomize when you want to see what the seed does.

loras accepts up to three weighted files. Feed it from a PopNow LoRA node: source takes a .safetensors sitting in ComfyUI/models/loras (uploaded once and cached for 12 hours, so a second run costs nothing extra), a Hugging Face repo id or page, a Civitai page, or a direct https:// link. Gated files read HF_TOKEN or CIVITAI_API_KEY from ComfyUI's environment. There's deliberately no token widget - widget values are saved inside every workflow and image you share, so a token there would leak with the first PNG.

No steps, no CFG, no negative prompt, no resolution tier. Turbo wants 8-9 steps at CFG 1 and the service runs its own schedule, so the practical loss is smaller than it looks.

How the run works

The node is generated from alibaba/z-image-turbo-lora/text-to-image's published schema, so every field is one the model actually accepts. On queue it resolves your LoRA sources (uploading local files, importing HF/Civitai references server-side), assembles the request, takes a quote first, holds at most that amount, submits with an idempotency key so retries can't create duplicate tasks, polls, then downloads and decodes the result into an IMAGE batch. You also get task id, which is what PopNow support will ask for.

Flat 1.8 credits per image - 1 credit = $0.001 - so roughly two cents a render with the LoRAs included. You can A/B a dozen prompt phrasings for the price of a coffee, which is exactly the loop this model is good at.

Install

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

Or Manager → search PopNow → install → restart. Then a key from popnow.ai/api/console/keys, entered under Settings > PopNow or exported as POPNOW_API_KEY, followed by one more restart so the node list refreshes from your account. No GPU, no model files, no extra dependencies.

Where people get burned

Stacking LoRAs on Turbo is the known soft spot: more than one at once tends to degrade the image, though people using LoKR report holding three at full strength. Start with one.

If your LoRA was trained on Z-Image Base, it'll look weak here - Base-trained LoRAs on Turbo want strength well above 1, and Turbo-trained ones don't transfer to Base at all. Check the base before you blame the node.

And a small one: with seed fixed and a prompt you keep editing, you're testing one variable at a time, which is what you want. Set it to randomize and you're testing two, which is how people conclude a LoRA "randomly stops working".