zImage / base / createImage

Z-Image Base is Alibaba's 6B Apache-2.0 model that the community crowned "the SDXL 2.0 we've been waiting for" - real seed-to-seed variety, negative-prompt support, day-one LoRA trainability, and a crisp render that's strong at anime and photorealism. This node runs it on Civitai's sdcpp fleet, so you get that without a 6B download and without the inference cost of the heavier 20B+ models. It's the well-rounded pick in the pack: cheaper than the Flux 2 nodes, smarter than SDXL.

The thing that makes this node stand out in the whole pack is its control_nets input - it's the only sdcpp image node in this batch that exposes one. Everything else in the Civitai image family is pure text-to-image; this one can take real structural conditioning. That's a meaningful differentiator if you want a cloud job to respect a pose or composition instead of guessing.

How it works

Standard pack client: submit engine: sdcpp, ecosystem: zImage, model: base, operation: createImage, long-poll, get a ComfyUI IMAGE back. The diffuser_model input is optional - unwired, the recipe runs the stock Z-Image Base.

The inputs that matter

• prompt (required) - natural language works well here, and so do tags. Z-Image Base was celebrated for actually varying across seeds, so iterate seeds for variety.
• control_nets - wire a Civitai ControlNet node (pick preprocessor, weight, step range, control image). Worth knowing: the community still treats Z-Image's ControlNet as the thin part of the ecosystem - usable for guidance, not as load-bearing as SDXL's. Don't expect miracles on the first try.
• diffuser_model - optional CIVITAI_AIR; leave empty for base, or wire a selector to try a fine-tune (Juggernaut and others now build on Z-Image).
• steps / cfg_scale - 20 / 4 defaults. Base wants more steps than the distilled Turbo variant; that's expected.
• negative_prompt - supported and effective; use it.
• loras - via a Civitai LoRA Selector. Z-Image stacks one LoRA well and gets flaky with several, so keep it lean.

Install

Pack install: ComfyUI Manager → search "Civitai Comfy Nodes" (publisher civitai) → Install → restart. Or clone https://github.com/civitai/civitai-comfy-nodes.git into custom_nodes, pip install -r civitai-comfy-nodes/requirements.txt (just requests), restart.

Common issues

Credentials: Civitai Auth node into api_config or CIVITAI_API_TOKEN (from civitai.com/user/account); headless installs need the env var.

The control-net caveat is the one to internalize: on the fleet, as on the desktop, Z-Image's ControlNet weights are early and weaker than the SDXL ecosystem's - if a job comes back that ignores the control image, it's the state of the art, not your wiring. Also, quantity goes to 12 and each render bills Buzz; seed-variety is free to explore, just not in bulk.