📥👑 Load Lora SDXL (+Download from CivitAi)

The SDXL LoRA version of Bjornulf's CivitAI loaders. You pick an SDXL LoRA from a curated in-node gallery (about 50 entries - detail tweakers, style LoRAs, the usual suspects), and the node downloads it from CivitAI if it's not already local, then applies it to your model and CLIP. Same as a Load LoRA node, but the dropdown is CivitAI and it hands you the trigger words too.

If you're the type who stacks a "add detail" or "better faces" LoRA on most SDXL gens, this cuts out the manual download step.

How it works

You feed it model and clip, choose a LoRA, and on run it checks locally; if the file's missing it pulls it from CivitAI into ComfyUI/models/loras/Bjornulf_civitAI/sdxl_1.0/, then patches the LoRA into your model and CLIP at your chosen strengths. Standard LoRA math - CivitAI is just the delivery mechanism. First download is a one-time cost; the file's local afterward.

The inputs and outputs that matter

• image - the LoRA picker, a thumbnail gallery of about 50 SDXL LoRAs. Your main choice.
• model and clip - from your SDXL checkpoint loader.
• strength_model and strength_clip - the LoRA's influence (default 1.0 each; range is a very wide −20 to 20 but sane values sit near 0.5–1.2). Lower strength_model if the LoRA is dominating the image.
• civitai_token - your CivitAI API token, needed only for the download.

Outputs: model and clip to chain onward, plus name, civitai_url, and trigger_words. Pipe trigger_words into your prompt so the LoRA actually activates.

How to install it

ComfyUI Manager: search Bjornulf_custom_nodes, install, restart. Manually:

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

Restart ComfyUI. The download depends on civitai-py from that requirements file. Windows portable: .\python.exe -m pip install ....

Common issues

The download half needs a valid token, and gated LoRAs can refuse regardless - check the token first if a first run stalls. Match the base model: an SDXL LoRA belongs on an SDXL checkpoint, not SD1.5 or Pony. Worth noting Pony counts as its own base - an SDXL LoRA will load on Pony but usually degrade the result, so use the Pony LoRA sibling for those.

If the LoRA appears to do nothing, add its trigger word (that's the trigger_words output) - a surprising number of "broken LoRA" reports are just a missing activation keyword. If it's overbaking the image, nudge strength_model down toward 0.6–0.8; SDXL detail and style LoRAs in particular are often trained hot and don't need the full 1.0.

The first download of any model is a real pause since SDXL LoRAs can run to a couple hundred megabytes each, so don't kill a run that looks stuck on first use. And keep the pack itself current - these CivitAI nodes had Windows encoding bugs that were fixed around v0.71–0.75. For other bases, use the SD1.5, Pony, or Hunyuan sibling nodes; this one lists SDXL LoRAs only.