Z-Image

Z-Image is the rare case where this pack wraps a model you can actually run yourself. Alibaba Tongyi-MAI's 6B image model is Apache 2.0, fits on a 12GB card, and generated the "Z-Image on 3060, 30 sec per gen" moment that killed Flux 2's launch hype. The community literally started calling it SDXL 2.0. So why use it through Runware? Because the model being small is exactly why the cloud version is cheap and instant - no 30-second local waits, no VRAM juggling, no scheduler fights, and the same quality if you're not married to local.

This node appears to track the distilled side of the family (the gated steps defaulting to 1 is a giveaway - the Turbo is guidance-distilled to a handful of steps). Either way, it's Z-Image quality behind an API, with the full editing toolkit the model is known for.

What you set

positivePrompt, width, and height are required (128–2048, step 16). The inputs that make this more than a T2I node:

• seedImage / maskImage - image + mask for inpainting. maskMargin (gated, 32–128px) adds context pixels around the masked region so the edit blends instead of smearing.
• outpaint - a RUNWARE_OUTPAINT builder socket for extending the canvas.
• lora - community LoRA socket; the Z-Image Turbo LoRA ecosystem exploded before the base even shipped.
• ultralytics - object-detection input, for workflows where you detect then generate.
• strength - gated, 0.8 default: lower = more of the original input survives.
• steps / scheduler - gated; with a distilled model, few steps is the point, not a compromise.
• seed - exposed, so local-familiar reproducibility carries over.

Output is image (IMAGE) → Preview/Save, outputFormat JPG/PNG/WEBP.

Installing

Part of the Runware/ComfyUI-Runware pack.

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

Deps are runware-sdk, pillow, soundfile. Runware API key from runware.ai/api-keys (Settings panel or RUNWARE_API_KEY).

Where people get burned

The honest question this node forces: why pay per image for a model that's free on your own GPU? The answer is "when you can't run it" - a 6GB laptop, a quick burst of batch work, or a workflow you want to hand off. If you own a 12GB card and use Z-Image daily, running it locally is still the better economics; this node is the convenience tier. And watch the usual cloud rules: everything you send goes to Runware's servers, and each numberResults is a separate billed generation.