🐈‍⬛ LoRA Metadata

LoRA Metadata looks like a normal LoRA loader, and it is - but the third output is the whole reason it exists. Beyond the usual MODEL and CLIP, it also hands you the LoRA's filename as a string, which you can feed into a display node, stitch into a prompt, or bake into the text you save with your image. It's a small quality-of-life trick, and for that one job it's genuinely handy.

One honest warning up front: the name oversells. "Metadata" here means "the filename," not the metadata block inside the safetensors file. This node will not show you trigger words, training settings, or CivitAI info - if that's what you want, that's rgthree's Power Lora Loader, which reads those off the file and fetches more from CivitAI. LoRA Metadata just labels the thing you loaded. Cheap, but it's a cheap trick done cleanly.

How it works

The node subclasses ComfyUI's built-in LoraLoader and rearranges it. The first change: model and clip move from required to optional. The second: it computes a clean filename (path stripped, and a trailing .safetensors dropped) and returns it alongside the patched model and clip.

That small change unlocks a useful mode. If you wire neither model nor clip in, the node doesn't apply anything - it just returns the filename and nothing else. So the same node doubles as a "pick a LoRA from a dropdown, get its name as text" picker. If you wire only one of the two, it applies the LoRA to that side and forces the other strength to 0 rather than erroring.

The one tiny quirk from the source: only .safetensors gets stripped from the filename. If your LoRA is a legacy .pt or .ckpt file, the extension rides along in the string output. Minor, but it'll bite you if you're matching filenames exactly.

Inputs and outputs that matter

• lora_name - a dropdown of every LoRA in your models/loras folder. This is the one you actually touch.
• strength_model / strength_clip - how hard to apply the LoRA, default 1.0, range −100 to 100. Both can go negative, which is occasionally useful for "un-applying" something. If you want a gentler look, 0.5–0.8 is the usual sweet spot.
• Outputs: model, clip (wire these where a normal LoraLoader's outputs go), and filename - a string you can feed into any STRING input, like the pack's own Display String node.

Installing it

Everything here is standard. In ComfyUI Manager, search "Super Loader" and install, or clone manually:

cd ComfyUI/custom_nodes
git clone https://github.com/markuryy/ComfyUI-SuperLoader

Then restart ComfyUI. The pyproject.toml lists no runtime dependencies, so there's nothing extra to pip-install and no model files to download. It's MIT-licensed and built from the official Comfy-Org cookiecutter template - note that the README is mostly that template's boilerplate about publishing and testing, which is why it reads like documentation for something bigger than it is.

Common gotchas

• LoRA not changing your output? You almost certainly left model or clip unconnected. Both are optional, and if both are empty the node does nothing on purpose. If you don't see the LoRA's effect, check you actually wired the MODEL in from your checkpoint loader.
• Expecting real metadata (trigger words, trained keywords) and finding only a filename - see above. That's this node's design, not a bug.
• The usual LoRA failure modes still apply regardless of loader: wrong base model family (an Illustrious LoRA won't work on Flux), missing trigger word, or weight too high. If a LoRA "doesn't work," test it with a fixed seed, with and without, before blaming the loader.