Z-Image Turbo Settings

Z-Image is Alibaba's 6.15B Apache-2.0 image model - the one that, per the KB's history, shipped into the same window as FLUX.2 Dev and took the local default because it was small, open and uncensored where the flagship was heavy and lobotomized. Z-Image Turbo is the distilled fast variant: 8 steps, CFG 1.0, done. AIOZImageTurboSettings is the pack's (deliberately small) control surface for it.

What's on it

The honest summary is: not much, and that's correct. The node emits one AIO_MODEL_SETTINGS output that goes into AIO Image Generate's model_settings socket with model_type set to z_image_turbo. The inputs are:

• force_steps - the exact sampling step count (default 8). This is the one knob with a real job. Z-Image Turbo is guidance-distilled: it expects a fixed low step count, and force_steps is how you pin it. Drop it to 6 and you get faster but rougher; 8 is the sweet spot the model was trained for.
• precision_policy - auto, fp8, bf16. Z-Image is small enough that you have options; auto picks a practical format for your runtime.
• The shared performance block: attention_mode (auto, sage, sage3, flash, xformers, pytorch, split, sub_quad, off), torch_compile_mode/torch_compile_backend (inductor or cudagraphs), and performance_apply_timing (after_loras vs before_loras).

That's it. No guidance (CFG 1.0 is the distilled contract), no sampler/scheduler pickers (the profile drives them as auto), and - notably - no prompt-enhancement or speed-preset controls, because the pack has no runtime implementation for them. The README says so in so many words rather than shipping dead UI. If you came from another Z-Image pack looking for a "quality/speed" preset dropdown, it's not here.

Wiring and install

Install the pack once and everything shows up under AIO/Image:

cd ComfyUI/custom_nodes
git clone https://github.com/helto4real/comfyui-all-on-one-image-generation-node
cd comfyui-all-on-one-image-generation-node
python -m pip install -r requirements.txt

Or via ComfyUI Manager (search "AIO Image Generate") and restart. Connect this node's AIO_MODEL_SETTINGS to model_settings on AIO Image Generate, set model_type=z_image_turbo, pick your Z-Image diffusion model, text encoder and VAE, and run. The profile defaults handle steps, CFG, sampler and scheduler.

Gotchas

• Negative prompts do nothing by default. The z_image_turbo profile reports supports_negative_prompt=False, and the main node's use_zero_negative_conditioning defaults to on. If a negative prompt matters to you, that's the toggle - but for Z-Image Turbo, it usually doesn't.
• No reference images, no inpaint. The README lists Z-Image reference and mask paths as staged for a later adapter pass. As of this version it's text-to-image only. Don't go hunting for sockets that don't exist.
• GGUF works if you bring the backend. The profile supports GGUF files, but only through an external backend like ComfyUI-GGUF - the pack won't load a .gguf on its own and will raise rather than silently fall back.
• It's a one-family node. Wire these settings into a Klein or Ideogram run and the main node rejects the mismatch. force_steps is Z-Image's, not anyone else's.

There's a certain virtue to a settings node this restrained. You set force_steps, maybe touch attention, and never think about it again - which is exactly the Z-Image experience.