Juggernaut Z

For two years, "Juggernaut" was shorthand for "realistic SDXL checkpoint" - the default answer whenever someone asked which model makes photos that look like photos. That SDXL line ended with Ragnarok in May 2025, and a year later the brand restarted on Z-Image as Juggernaut Z, which is what this node runs (rundiffusion:200@100). If you knew Juggernaut from the CivitAI era, this is the same mission on a modern base.

Z-Image is Alibaba's Apache-2.0 image model, and the Juggernaut treatment - a curated fine-tune tuned for photorealism - is exactly what that base needed to win back the realism crowd. Via Runware you don't download any of it; the node points at the hosted checkpoint, and your job is prompt, size, and the sampler knobs.

How it works

imageInference with the rundiffusion:200@100 AIR. Standard Runware mechanics: prompt and any IMAGE inputs go up, a rendered image comes back as a native IMAGE tensor. The node exposes the full SDXL-era control set because the checkpoint still behaves like the lineage it came from: steps (default 35), scheduler (75 choices, defaulting to a flow-match scheduler - that's the Z-Image world, not DPM++), CFGScale (default 6), negativePrompt, seed, and strength for img2img.

The inputs that matter

• positivePrompt (required) - this is a realism model; describe scenes, materials, and light like you'd photograph them.
• width / height - required, 1024×1024 defaults, 128–2048 range in steps of 16.
• negativePrompt - still does real work here. "blurry, oversaturated, plastic skin, bad anatomy" is the classic list.
• CFGScale - 6 default. Juggernaut has always wanted a fairly high CFG for punch; if faces look mushy, don't touch CFG, add steps.
• seedImage / maskImage - the img2img and inpainting inputs, with maskMargin for extra context around the masked area.
• lora - the RUNWARE_LORA socket; Juggernaut's LoRA ecosystem is exactly why people reach for it.
• steps - 35 default; it's a Z-Image base so it's happier at lower step counts than old SDXL, but the default is a safe floor.

Installing it

Install the pack, not the checkpoint. ComfyUI Manager → search Runware → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Key from runware.ai/api-keys → ComfyUI Settings → Runware API key (or RUNWARE_API_KEY). The Juggernaut Z checkpoint is hosted - no download.

Where people get burned

• It's a hosted fine-tune, so version matters. rundiffusion:200@100 is one pinned checkpoint AIR. If RunDiffusion ships a newer Juggernaut Z, this node keeps pointing at the old one until the pack updates - which is actually the safe behavior.
• The Z-Image base changed the sampler defaults. Don't carry over SDXL muscle memory: the flow-match scheduler and lower step counts are deliberate. If your old Juggernaut recipe overdevelops, that's why.
• Cost per run. Cloud inference is metered and the title bar shows each run's price. Juggernaut's higher default steps are a quality choice and a price choice.
• Don't assume uncensored. Juggernaut's SDXL line was famously "nudity yes, porn no," and the platform split the brand's SFW/unrestricted editions - the hosted version is the one Runware serves, so moderation behavior rides along with it.

Juggernaut Z is the closest thing to "the old default, modernized" in this pack. If you lived in the Juggernaut era, you'll feel at home; if you didn't, it's a strong place to start for realistic output.