Stable Diffusion XL (custom model)

This is the node that turns "which checkpoint should I use?" from a download decision into a search box. RunwareArch_sdxl is a custom-model node: it hosts the whole SDXL architecture, and the model field takes any SDXL checkpoint's AIR - civitai:101055@128078 by default - so you can swap community fine-tunes without downloading a single weight file. It's Runware's answer to the fact that the SDXL ecosystem is thousands of checkpoints and your hard drive is not.

If that sounds niche, think about who this is for. You don't have a GPU, or you're on a laptop, or you just don't want to manage 40GB of SDXL variants. You browse the model catalog (the node has a search button), pick a checkpoint, and it runs in the cloud like it was local. The SDXL fine-tune ecosystem - anime, realism, illustration - is all reachable through one node, and the sockets for LoRA, ControlNet, IP-Adapter, embeddings, refiner, and PhotoMaker are right there to stack on top.

How it works

The node sends an imageInference request with whatever AIR is in the model field. Your IMAGE inputs (seed image, mask) go up as base64, the sampling parameters you set are passed through, and the finished image returns as a native IMAGE tensor. Every feature socket - lora, controlNet, ipAdapters, embeddings, refiner, photoMaker - is a typed input that wires to the pack's builder nodes, which is how you chain, say, a ControlNet edge map plus a LoRA in one request.

The inputs that matter

• model (required) - the AIR string, default civitai:101055@128078. Use the search button to browse the catalog instead of typing from memory.
• positivePrompt (required) - plus negativePrompt, which SDXL still rewards.
• width / height - 1024×1024 defaults, 128–2048, step 8. SDXL likes 1024-era resolutions; straying far degrades anatomy.
• steps / CFGScale - interesting twist: both are off by default, gated behind a checkbox each. The node deliberately lets the model decide unless you intervene. Enable and set steps_value / CFGScale_value when you want control.
• seedImage / maskImage - img2img and inpainting, with maskMargin for context pixels.
• clipSkip and vae - the classic SDXL tinkering pair; vae overrides the checkpoint's bundled VAE with an AIR.
• advanced_json - documented escape hatch for hiresFix.

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).

Where people get burned

• Every checkpoint is a different bill. Cloud inference prices by compute; a heavy 50-step run on a big SDXL merge costs more than a quick one. The title bar shows each run's price - that's your feedback loop.
• Not every checkpoint is on Runware. The search button shows the catalog; an AIR that 404s means that specific CivitAI file isn't hosted. Pick from what's there.
• The gated defaults are a feature. steps and CFGScale being off by default means the request goes out without them and the model fills in sensible values. If you set them, set them on purpose - a 0 CFGScale_value with the gate on sends guidance that may surprise you.
• Feature sockets need builders. controlNet doesn't accept a raw image; it takes the pack's ControlNet builder output (like the Canny preprocess node's edge map).

RunwareArch_sdxl is the node that finally makes "try twenty SDXL checkpoints" a twenty-minute experiment instead of a twenty-gigabyte download spree. It's the most flexible image node in the pack.