Z Image Turbo (custom model)

If you've already met the dedicated Z-Image-Turbo node, this is its sibling with a twist: RunwareArch_z_image_turbo is the custom-model version, so instead of being pinned to the stock checkpoint, the model field is yours to fill with any Z-Image-Turbo-architecture model from the Runware catalog. Same 6B family that became the community's SDXL 2.0 - near-FLUX quality, tiny, uncensored, Apache 2.0 - but now you can point it at the specific checkpoint you actually want.

Why would you reach for the Arch version over the fixed node? Fine-tunes, mostly. Z-Image built a LoRA and fine-tune ecosystem faster than almost any model in recent memory, and the dedicated node is stuck with the base. This one lets you load a community Z-Image checkpoint or a merged variant and get its exact behavior in the cloud - no 3GB download, no VRAM negotiation. It's the same "doorway node" pattern as the pack's other RunwareArch_* entries, applied to the hottest small model of the year.

What you set

Required: model (default runware:z-image@turbo - search button to browse), positivePrompt, width, height (128–2048, step 16).

• CFGScale toggle + CFGScale_value - Z-Image-Turbo is guidance-distilled, so the same advice as the fixed node: keep it low, usually leave it off.
• steps - the Turbo architecture wants few steps; the default is the distillation-correct value.
• controlNet, lora, outpaint sockets - Z-Image's real ControlNet ecosystem (Alibaba's Fun ControlNet) plugs in here.
• seedImage, maskImage, negativePrompt (ignored at CFGScale ≤ 1), scheduler, seed, numberResults (up to 20), outputFormat, outputQuality.

Output is image (IMAGE).

How it works

Standard pack machinery: request with the model AIR + taskType over REST via the Runware SDK, IMAGE downloaded back, cost on the title bar. Because the architecture is known to the catalog, sampler handling is automatic - and for a distilled model like Turbo, "automatic" is doing a lot of work, because the correct few-steps/low-CFG settings are exactly what beginners get wrong locally. The cloud also means the 12GB-VRAM sweet spot that made Z-Image famous is irrelevant here; a potato laptop drives this node identically.

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 via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

Search, don't type the AIR - a stale or mistyped model string is the classic request failure. If a fine-tune behaves differently than its page promised, remember cloud runs may use the catalog's sampler defaults; nudge CFGScale and steps to the values the checkpoint page recommends. And the negative-prompt-at-low-CFG thing applies here too, so don't fight it.