Stable Diffusion XL Lightning (custom model)

SDXL Lightning was ByteDance's February 2024 answer to the speed problem: take SDXL, distill it so it renders in a handful of steps instead of thirty, and ship it under a name that promises exactly what it delivers. RunwareArch_sdxl_lightning is the custom-model node for that architecture - the model field defaults to a Lightning checkpoint (civitai:350352@391971) but takes any Lightning-style SDXL AIR - so you get the fast SDXL family in the cloud, no downloads, no VRAM.

The distillation context is the whole reason this node exists. Distilled models trade detail for speed - the KB's line is that every distillation "drastically reduces" some quality - and Lightning's design assumption is 1–4 steps with CFG near or at 1. In Runware's cloud, that collapse of steps is a collapse of cost per run, which makes this the budget SDXL node: same checkpoint ecosystem, dramatically cheaper iterations. It's the right rung for drafts, batches, and anything where "good enough at a tenth of the price" wins.

How it works

imageInference with the AIR in the model field. Standard mechanics - prompt and IMAGE inputs up, image back. The sampling controls are the full SDXL set (scheduler with 76 choices, gated steps/CFGScale, clipSkip, vae, promptWeighting) plus the feature sockets (lora, controlNet, ipAdapters, embeddings, refiner). Note both steps and CFGScale are off by default - for a distilled model, the checkpoint's own low-step defaults are usually the right call, and overriding them with SDXL-era habits (30 steps, CFG 7) is how you get oversaturated, overcooked output.

The inputs that matter

• model (required) - the AIR, default civitai:350352@391971. Search the catalog for other Lightning checkpoints.
• positivePrompt (required) - plus negativePrompt; Lightning still honors a good negative list.
• width / height - 1024×1024 defaults, 128–2048, step 8.
• steps / CFGScale - gated. When you enable them, think "1–4 steps, CFG near 1," not "SDXL normal." The distillation doc is blunt: the settings are completely different from the standard version.
• seedImage / maskImage - img2img and inpainting, with maskMargin.
• lora / controlNet / ipAdapters - the stackable feature sockets.

Installing it

Install the pack. 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

• Bringing SDXL settings to a Lightning model. 30 steps and CFG 7 on a 4-step model is the classic recipe for garbage. Trust the low-step defaults; that's the distillation.
• "Fast" has a ceiling. Lightning output is draft-grade versus full SDXL. For final work the pack's regular RunwareArch_sdxl or Runware_rundiffusion_juggernaut_z are the higher rungs.
• Cost is lower, not zero. Each run is still metered; the title bar shows the price. It's just a much smaller number than full SDXL.
• Checkpoint availability. The catalog is what it is - the search button shows hosted AIRs, and an unserved ID 404s.

SDXL Lightning custom model is the "cheap and quick SDXL" node. When you need to see twenty variations of an idea before committing a single full-price render, this is the rung.