Grid LoRA Loader

Scrolling a plain dropdown of forty LoRA filenames trying to remember which one made that anime-style render you liked last week is nobody's idea of a good time. Grid LoRA Loader replaces the dropdown with an actual thumbnail browser - pictures instead of filenames - plus the same per-entry strength and trigger-word handling you'd want from a serious multi-LoRA setup.

What it is and where it fits

This node, and its list-view sibling List LoRA Loader, are CrasH Utils's take on a pattern the ecosystem already knows well from rgthree's Power Lora Loader - a single node that stacks multiple LoRAs with per-entry toggles instead of chaining loader nodes end to end. The CrasH Utils README credits that pack directly as the UI inspiration. What sets Grid LoRA Loader apart isn't the stacking (that part's table stakes at this point) - it's the visual browser: a thumbnail grid, per-folder navigation, search, and an NSFW-folder filter, none of which rgthree's version does.

How it works

The node has three panels you can resize by dragging the dividers between them. Browse lets you navigate folders, search, toggle the SFW filter, and double-click a thumbnail to add or remove it from your active stack (right-click for info). Selected shows what's currently active - single-click to enable/disable, double-click to remove, the gear icon to adjust strength. Trigger Tags lets you pick which trigger words actually get sent downstream.

Thumbnails come from an image sitting next to the .safetensors file with the same base name (.png preferred if multiple formats exist). If you don't have one yet, generate an image with that LoRA, right-click the Preview Image output, and pick "Save as LoRA thumbnail" - it saves as .png and overwrites anything already there. Trigger words work the same way as the list variant: a companion .txt file, same base name, one word per line or comma-separated.

The SFW toggle hides any folder named NSFW (case-insensitive) at any depth in your library - flip it off if those folders should be visible again.

Inputs and outputs that matter

Both real inputs, model and clip, are optional - but in practice you'll always wire in your checkpoint's model and CLIP, since that's what actually gets patched by whatever LoRAs you've selected.

Outputs: MODEL and CLIP, the patched versions ready for your sampler and text encoder, plus triggers - a comma-separated STRING of whichever tags are toggled on in the Trigger Tags panel, ready to feed into your positive prompt.

Installing it

ComfyUI Manager: search "CrasH Utils". Manual:

cd ComfyUI/custom_nodes/
git clone https://github.com/chrish-slingshot/CrasHUtils.git

Restart ComfyUI. No extra dependencies.

Common issues

A LoRA has no thumbnail. Either drop a matching-basename image next to the .safetensors yourself, or generate something with it loaded and use the right-click "Save as LoRA thumbnail" option on a Preview Image node - it's the fastest way to build up a visual library over time.

Folders missing from the browser entirely. Check the SFW toggle - any folder with "NSFW" in the name gets hidden at any depth, and it's easy to forget you've got it switched on.

Trigger words not populating. The companion .txt needs the exact same base filename as the LoRA and needs to be right next to it - a mismatched name means the loader silently finds nothing.

LoRA loads but has no visible effect. Standard LoRA checklist applies here same as anywhere else: confirm base-model compatibility (an SDXL LoRA won't do anything on Flux), that the trigger word is actually in your prompt, and that the strength isn't set somewhere it shouldn't be for that particular LoRA - there's no universal "correct" weight across architectures.