Juggernaut Pro Flux

Juggernaut is the realism checkpoint with a fan club. The SDXL-era run of it became the default answer to "which model actually looks photographic" for years, and the FLUX line - Juggernaut Pro Flux - was the family's leap to the new architecture. If you've been around since the Juggernaut XL days, this node is that lineage, running server-side on Runware.

That matters more than it sounds, because the original "Juggernaut FLUX Pro vs FLUX Dev" release was API-only. You couldn't just drop it into a local ComfyUI like you could the SDXL weights. This node is that same arrangement made convenient: RunDiffusion publishes the model on Runware's platform, and Runware's pack turns it into a graph node with a real IMAGE output. No checkpoint file to hunt down, no models/checkpoints folder, no 20GB download. You pay per image instead.

What you actually set

positivePrompt, width, and height are required; defaults land you at 1024×1024 with a 64px step. Everything else worth touching:

• steps / steps_value - the enable-then-set pattern this pack uses everywhere: flip steps on and steps_value appears. FLUX-family models are happiest around 20–30 steps. Leaving it off just uses the model's default, which is a genuinely fine place to start.
• CFGScale_value - guidance. For FLUX derivatives, keep it low (1–4 range); the tooltip's reminder that negativePrompt is ignored when guidance is off (CFG ≤ 1) is worth taking literally.
• scheduler - a 76-entry dropdown covering every scheduler anyone has ever shipped. Leave it at (default) until you know why you're changing it.
• seedImage + maskImage + outpaint - the img2img/inpaint path. maskImage takes a native MASK, and maskMargin_value (32–128px) tells the model how much context to see around the edit so the inpaint blends instead of floating.
• lora, controlNet, ipAdapters, pulid - typed sockets that accept the pack's builder nodes (Runware/Params). This is where Juggernaut shines: it's a character/style-friendly model, so a LoRA or PuLID identity feeding in here is a legitimate workflow.

strength governs how hard the source image is transformed; lower = more of the original survives. numberResults goes up to 20, so it's cheap to farm variations. advanced_json is the escape hatch if you want hiresFix or any param the widget list doesn't expose.

How it runs

Under the hood this node builds a request (model: rundiffusion:130@100, task imageInference), ships it over the Runware SDK's REST transport, and converts the returned image URL into a native IMAGE tensor. Because it's cloud inference, the result flows into Preview Image or Save Image exactly like a local generation - and the node's title bar shows the run's cost and content-check verdict. That's a nice feedback loop when you're burning through variations.

Install and gotchas

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

Restart, then grab an API key from the Runware dashboard and paste it into ComfyUI Settings → "Runware API key" (or set RUNWARE_API_KEY, or runware auth login). It's a paid API - minimum top-up, card required, cost printed per run.

Where people get burned: they treat it like the local FLUX habit and crank CFGScale to 7. Juggernaut Pro Flux is a fine-tune, not base FLUX dev - the tuning already did the prompt-following work, so high guidance gets you burned colors and mush. Start at the model defaults, move one knob at a time, and let the per-run cost on the title bar keep you honest.