Z-Image-Turbo

Z-Image-Turbo is Alibaba's 6B image model, and it's the one that quietly won the community's heart in late 2025 by being small at exactly the right moment. While FLUX 2 Dev shipped 32B parameters and a heavy censorship load, Z-Image shipped a tiny, fast, uncensored, Apache-2.0 model that runs on a 12GB card - near-FLUX quality at a fraction of the size, and the community called it SDXL 2.0 before it even had a trainable base. In the Runware pack it's the sensible middle ground: better detail than the distilled tiers, cheaper per image than the 20B flagships, and Turbo specifically means it does it in very few steps.

It's not perfect. Complex-prompt adherence is weaker than FLUX 2, there's a persistent Asian face bias, and the native ceiling is around 2MP - you won't be printing billboards. But for fast, good-looking, no-drama generation, this is the node people keep coming back to.

What you set

Required: positivePrompt, width, height (128–2048, step 16).

Then it's the familiar Runware image-node spread:

• CFGScale - off by default and leave it off. Z-Image-Turbo is guidance-distilled; the whole point is low CFG at few steps.
• steps - the default handles it. Adding steps on a Turbo model is throwing money at a ceiling.
• seedImage / maskImage - image-to-image and inpainting, standard.
• controlNet, lora, outpaint sockets - Z-Image has a real ControlNet ecosystem (Alibaba's Fun ControlNet), so this is one of the more expandable models in the pack.
• negativePrompt (ignored at CFGScale ≤ 1), seed, scheduler, numberResults (up to 20), outputFormat, outputQuality.

Output is image (IMAGE).

How it works

Same pack machinery as every node here: build a request with the model AIR and taskType, ship it over REST through the Runware SDK, download the IMAGE back. The model runs on Runware's hardware, so the 12GB-VRAM sweet spot that made Z-Image famous locally doesn't even matter - you could run this on a Raspberry Pi's worth of client and get identical results. The node prints cost per run on its title bar; Z-Image-Turbo's cost is low enough that batch experimentation stops feeling like a budget decision.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

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

No model downloads; dependencies are runware-sdk, pillow, soundfile. API key needed: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

The missing-key error is the usual first stop. If you came from SDXL habits and set CFG 7 and 30 steps, you'll get overcooked output - trust the defaults, they're the distillation-correct values. And if you need ultra-detailed photorealism, Z-Image-Turbo isn't lying to you about being a turbo; the base model (or the FLUX nodes) is the step up.