Z-Image Turbo (Replicate)

Z-Image Turbo is the odd one out in the image half of this pack: unlike Seedream or Nano Banana, it's an open model. Alibaba Tongyi's 6B turbo-distilled image generator, released alongside the Nano Banana Pro launch week in late 2025 and immediately adopted as the open answer. So the honest question you should ask before using this node is: why rent it?

The answer is "no GPU, or not enough of one." Z-Image Turbo is a 6B DiT - genuinely runnable locally on a mid-range card, and at 4–8 steps it's fast. But if your ComfyUI lives in the cloud or on a machine without a serious GPU, this node gives you the same model on Replicate's hardware with zero setup. That's the legitimate reason-3 use case from the API-node playbook: open model, someone else's GPU, pay per call.

How it works

prunaai/z-image-turbo from the pack catalog, standard engine. Prompt goes up, the API runs its distilled denoiser at your step count, the image comes back as an IMAGE tensor. Nothing about the cloud round-trip changes the model's behavior - you just don't own the silicon.

The inputs that matter

• prompt (required) - text prompt. This is a turbo model, so it follows prompt intent fast and is happiest with clear, specific descriptions.
• width / height - 64–2048, both defaulting to 1024. This is a 1024-native model; going far past that invites artifacts.
• num_inference_steps - default 8. This is the turbo dial and the one setting that actually changes speed/quality balance. 8 is the sweet spot; 4 is draft-tier; past ~12 you're burning money for nothing.
• guidance_scale - default 0, and the tooltip is emphatic: should be 0 for Turbo models. This is the distilled-model trap from the distillation playbook - people crank CFG out of habit and wreck the output. Leave it at zero.
• seed - set for reproducibility.
• go_fast - off by default; additional optimizations for speed. On the paid tier it's a cost lever: flip it on for iterations, off for finals.
• output_format / output_quality - jpg at 80 by default. png if you're editing downstream.

Single output: image, an IMAGE tensor, straight into SaveImage.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", restart. Dependency: replicate>=1.0.7. Token as an environment variable:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• Cranking guidance_scale. The single most common mistake with any turbo model. Zero is a feature, not a bug - the model was distilled with CFG disabled.
• Running steps you don't need. 8 is the target; every extra step is latency and cents with almost no quality return.
• Renting what you could run. If you have a GPU that can handle a 6B DiT, weigh the local path - you'll iterate faster and free than per-call will ever be. This node shines exactly when that's not an option for you.