Zhenzhen_Z_Image_Turbo

Alibaba's Z-Image is the less-famous sibling in the same Tongyi family as Qwen-Image, and the turbo variant is exactly what it sounds like: a fast, cheap image model tuned to generate in a handful of steps. Comfly_Z_image_turbo is the ComfyUI relay for it - prompt in, image out, default 8 inference steps, and the README's framing is basically "fast and cheap, added 20251205." It's the node you reach for when you want to iterate quickly and don't need the heavier Qwen generation model's polish.

Like everything in Comfyui_Comfly it's a paid call to ai.comfly.chat: POST to the relay's /v1/images/generations endpoint with the z-image-turbo model, decode the returned image, hand you a tensor. No local weights, no GPU, API key from the same site.

Inputs

• prompt - the only thing that really matters.
• size - seven presets from square 512 to cinematic 1536x1024, plus Custom, in which case custom_size takes a literal 1024x1024 string.
• output_format - jpeg (default), png, or webp.
• num_inference_steps - default 8, the turbo signature. You can raise it toward the max of 50, but at that point you're fighting the model's whole reason for existing.
• guidance_scale - default 0.0 here, which is a real Z-Image quirk: this model family is happy with no guidance, and cranking it isn't the win it is in SD. Tweak only if you have a specific problem.
• output_quality - 0–100, default 80; matters for jpeg/webp.
• seed - reproducibility.
• apikey - optional, as always.

Outputs

image - a real IMAGE tensor, wire it to a Save Image node. image_url - the hosted file link. response - the formatted generation log, and the place errors appear (a bad custom_size format, for instance, returns a white image plus the message here).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ainewsto/Comfyui_Comfly

or ComfyUI Manager → search "Comfyui_Comfly" → install → restart. API key from ai.comfly.chat into the apikey field, saved to the pack's Comflyapi.json for the rest of the pack.

The honest take

This is the pack's "draft" node. When you're exploring a concept and want ten variations without watching a progress bar or a meter climb, 8-step turbo is the right tool, and at the pack's pricing the per-image cost is close to negligible. Where it's not the right tool: anything where text rendering or fine detail matters - that's Comfly_qwen_image's lane. One note if you've used other Z-Image tooling: the KB's Qwen-Image coverage doesn't extend to a dedicated Z-Image panel yet, so treat the turbo model as "familiar family, lighter weight, lower expectations" and judge it on its own terms. For roughs, mockups, and seed-of-an-idea hunting, it's hard to beat on price.