Z Image (custom model)

If you've been in the ComfyUI ecosystem for more than a few months, you know Z-Image: Alibaba's open-weights model that quietly took the photorealism default away from Flux. This node is that model plus one very good twist - it's an architecture node. Instead of being locked to the base Z-Image, it runs any checkpoint of Z-Image architecture from the Runware catalog. The model field takes an AIR string (the default is runware:z-image@0) and there's a search button right on the node to browse community fine-tunes. Same node, different model, no re-installing anything.

This is the payoff of how the Runware pack is built: one node per model, generated from a schema, all running on Runware's cloud. You install the pack once (ComfyUI Manager → search Runware, or clone https://github.com/Runware/ComfyUI-Runware into custom_nodes/), add an API key in ComfyUI Settings, and every model - including every Z-Image fine-tune they host - shows up as a node. No checkpoints, no VRAM math, no models/unet folder to babysit.

How it works

Mechanically it's the same as every image node in the pack: your widgets and any wired builder sockets get folded into a request, sent to Runware as a taskType of image inference, and the result comes back as a native IMAGE tensor. The difference is the model input, which the node reads per-run. Swap runware:z-image@0 for a fine-tune's AIR and the same graph renders through that checkpoint. That's the whole value proposition: a workflow that uses this node becomes a testbed for every Z-Image-architecture model in the catalog.

Inputs that matter

• model (required) - the AIR, with a search button. This is the field you'll actually change.
• positivePrompt - Z-Image is a T5/LLM-encoded architecture; write descriptive natural language, not SD1.5 tag soup.
• width / height - 128–2048, stepped by 16.
• seedImage / maskImage - IMAGE sockets for img2img and inpainting; pair with strength and maskMargin (the inpainting context-padding control).
• CFGScale - here's the subtle one. Z-Image is guidance-distilled, and the tooltip is blunt: the negative prompt is ignored when guidance is disabled (CFGScale ≤ 1). The base model wants CFG 1 - that's the default state of the art for 2026 models, guidance baked into the weights. Cranking CFG on Z-Image doesn't buy prompt adherence; it overcooks the image.
• steps - toggle-gated; off uses the model's default, which is usually the right call on a distilled architecture.
• lora / outpaint / acceleratorOptions / watermark / ultralytics - typed sockets from the Runware/Params builders.

The image output is a plain IMAGE, so it drops into Preview Image or Save Image exactly like a local generation. The title bar shows your per-run cost and the NSFW flag when a content check ran - useful, since a custom fine-tune may behave differently from the base.

Where people get burned

• CFG ≠ more adherence here. The "CFG 7 fixes everything" muscle memory from the SDXL era actively hurts on Z-Image. Leave the gate off, or stay in the 1–3 range.
• Fine-tune AIRs - the search button is your friend; typing a random AIR gets you an error with a parameter/code tag that tells you what to fix, but it's still a wasted run.
• The toggle pattern - steps/CFGScale/strength are "enable to set X" booleans with paired _value fields. Flip the toggle, set the value; off means model default.

The name is a little dry for what it is. This is the node that turns "try the new Z-Image fine-tune" from a half-hour of downloading into a dropdown change.