Stable Diffusion XL Turbo (custom model)

SDXL Turbo was the model that made "how fast can you go?" a real question. It's the Adversarial Diffusion Distillation take on SDXL - a student model trained to jump straight to the answer in a handful of steps instead of 20–50. The tradeoff is baked in: fewer steps means less refinement, and distilled models are picky about settings (low CFG, the right scheduler, don't push it to high resolution). But for iteration speed, it's still hard to beat.

This node is the Runware pack's architecture wrapper for it. The model field defaults to the actual SDXL Turbo checkpoint (civitai:215478@242846) but it's a "custom model" node, so you can swap in any SDXL-architecture checkpoint from Runware's catalog and run it without downloading anything. The magic of this particular arch node is how much of the SDXL toolchain it exposes.

How it works

Standard pack mechanics: the node assembles a taskType: imageInference request with your model AIR and widget values, ships it over the Runware REST API, and converts the returned image URL into an IMAGE tensor. Nothing runs on your machine except the request.

Inputs that matter

• model - the AIR. Leave the default unless you want a different SDXL checkpoint.
• positivePrompt - plus negativePrompt, which only applies when CFG is above 1.
• seedImage / maskImage - wire an image in and you've got img2img / inpainting. maskMargin adds context pixels around the mask (32–128) so the inpainted area blends instead of looking pasted.
• lora / controlNet / ipAdapters / embeddings / refiner - typed sockets for the pack's builder nodes. This is where the node shines: ControlNet for pose/structure, IP-Adapter for style/character, LoRA for whatever the checkpoint is known for. Stack them.
• strength - img2img strength, off by default (this model has usage rules, so it's gated behind an enable toggle).
• steps / scheduler / CFGScale - with a distilled model, less is more. A handful of steps, CFG near 1, and the model's preferred scheduler beat brute-forcing 30 steps at CFG 7.
• advanced_json - the escape hatch; for this node it's documented for hiresFix. Paste a JSON object and it merges into the request.

Install

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

Or ComfyUI Manager → search Runware → Install → restart. API key in ComfyUI Settings → Runware API key or the RUNWARE_API_KEY env var. The pack is prepaid top-up; the node title bar shows cost per run.

Gotchas

• Distilled models punish you for treating them like normal ones. If your Turbo output looks flat, lower CFG, cut steps, and check the scheduler - don't just add steps.
• JPG output flattens alpha, which matters if your inpaint workflow wants masks later.
• If a run errors, the console message includes the offending parameter - the pack deliberately surfaces parameter 'x' so you know which widget broke the request.

The reason to pick this over a generic SDXL node: it bundles the whole SDXL accessory kit - ControlNet, IP-Adapter, refiner, inpaint - behind one remote call. Fast, cheap, and surprisingly full-featured for something that never touches your GPU.