zImage / turbo / createImage

Here's the irony you should sit with before using this node: Z-Image Turbo is one of the best local models to come out in years. Alibaba's Tongyi-MAI released it open (Apache 2.0), it's only 6B parameters, it runs on a 3060 in about 30 seconds, and the community went genuinely wild for it - "Z-Image on 3060, 30 sec per gen" was the +1,958 thread. You can absolutely run it yourself. So why would you run it on Civitai's servers for Buzz?

Because you might not have a 3060. Or a card at all. This node puts Z-Image Turbo inside ComfyUI when you're on a laptop, a Mac, or a rented box where installing torch isn't happening. It's the "the model is free, my GPU is not" case, and it's a legitimate one - you're paying for compute, not access.

How it works

The node is a recipe wrapper: it sends your prompt plus settings to Civitai's Orchestration API under the sdcpp engine discriminator (that's the Sdcpp in the class name - the server-side inference stack Civitai uses for this ecosystem), polls the workflow, and returns the image as a native IMAGE tensor. The steps default of 9 and cfg_scale default of 1.0 are exactly right, because Z-Image Turbo is a distilled model - guidance-distilled, so it wants CFG 1 and 8-ish steps. If you've read the Z-Image lore, those defaults are the distilled reality; bumping steps to 30 does not make it better.

Inputs and outputs

Only prompt is required. The interesting optional ones:

• diffuser_model - a CIVITAI_AIR socket. Wire a Civitai Model Selector here to swap which checkpoint the job runs on. This is the node's "which model" knob, since Z-Image has fine-tunes on Civitai.
• loras / control_nets - the selector-based stacks from the Loaders menu. Yes, this cloud node takes LoRAs and ControlNets, which is the whole reason to use it over the plain web generator.
• sample_method / schedule - the full sdcpp sampler list (euler, dpm++, euler_a, ... and schedules karras, ays, etc.). Leave them empty unless you know what you're doing; empty means "server default."
• negative_prompt, steps, cfg_scale, seed, quantity (up to 12), width/height (64–2048).

Outputs: images (IMAGE - wire to a preview or Save Image), plus errors, workflow_id, raw_json for debugging and cost inspection.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Or via ComfyUI Manager (search "Civitai Comfy Nodes"). Then authenticate - CIVITAI_API_TOKEN env var for headless, or the sidebar connect panel. No model download needed; the fleet has the weights.

Gotchas

Two things. First, the same model locally is free; on the farm it costs Buzz per image, and queue time on Civitai's side means "slow" is normal for this whole pack - the launch thread's main complaint was exactly that. Second, this is an early-preview pack generated from the orchestration spec, so the sampler list and defaults can change between releases without ceremony.

The trap nobody mentions: if you have a working local setup, this node is usually a downgrade that costs money. Reach for it when the machine in front of you can't run the model - that's what it's for. When it can, run Z-Image Turbo yourself and keep the Buzz.