Neons LoRA Explorer

You download soft_light_v3.safetensors and that filename is the only information you will ever get. Does it want a trigger word? Is it too strong at 1.0? Does it flatten everything into the same face on your checkpoint? Every LoRA loader in ComfyUI answers none of that - they're all dropdown, two numbers, done.

Neons LoRA Explorer is the same loader - MODEL and CLIP in, MODEL and CLIP out - with one idea bolted on: a preview gallery beside the node, filed per model folder. Same LoRA tested on two checkpoints, two separate piles of images, because it does not behave the same way on both.

Why this instead of what you already have

If your problem is "I have 400 LoRAs and can't find anything," you want a manager - LoRA Manager is the community default, a 332-mention topic across 2025–2026. If you want to stack five LoRAs with toggles, rgthree's Power Lora Loader does that and can pull trained words off a file or CivitAI.

This node is narrower: a single-LoRA loader whose selling point is visual memory, and it works because it needs no database - your folder structure is the database. Nothing is automatic, though. No crawl, no auto-populate; press Save or get no thumbnails.

How it actually works

The dropdown is ComfyUI's own list, straight out of folder_paths.get_filename_list("loras") - whatever ComfyUI can see is what you get, cached for about 20 seconds and refreshable without a restart via Rescan. The path is then split into three pieces:

loras/krea 2/film_grain.safetensors            gallery "krea 2"
loras/krea 2/portraits/soft_light.safetensors  gallery "krea 2", family "portraits"
loras/loose_one.safetensors                    gallery "Unsorted"

Top-level folder is the gallery, the folder inside it is a family, loose files land in "Unsorted". No config, no tagging step - you sort it by moving files around.

Applying is what the core loader does: read with comfy.utils.load_torch_file(path, safe_load=True), hand it to comfy.sd.load_lora_for_models(model, clip, weights, strength_model, strength_clip). Previews land in user/loras/<gallery>/, trigger words in user/loras/triggers.json, and your loras folder is never written to.

The inputs and outputs that matter

Required: model and clip from your checkpoint loader, lora (the dropdown, default None), and strength_model / strength_clip, both 1.0 by default and both accepting −20 to 20. Negative values invert the LoRA - a real trick for subtracting a concept, not a typo guard.

In practice you set two things: the LoRA, and strength_clip separately from strength_model - on some LoRAs the text-encoder half is what overcooks adherence while the model half is fine. And the usual failures still apply: wrong base (an Illustrious LoRA on Flux loads and does nothing), missing trigger word, or 1.0 simply being too strong. Try 0.5–0.8 before you blame the file.

Four outputs. model and clip go into your sampler and text encoders. lora_name is a STRING holding the LoRA's path - handy as a Save Image filename prefix so you can tell later which test render used which adapter. triggers is a STRING you type yourself under the preview; wire it into a text concatenation and the words travel with the LoRA instead of living in your head.

One sharp edge: zero both strengths and the node hands your inputs back untouched, bypassing silently - which is what you asked for.

Installing it

ComfyUI Manager, search Neons Style Explorer, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Neon-Sparks/ComfyUI-NeonsStyleExplorer

Then restart. No requirements.txt and no install script - Pillow and numpy ship with ComfyUI, and the load path is ComfyUI's own comfy.sd. No models to download, MIT licensed, and the pack's 3015-entry style catalog comes along whether you wanted it or not.

The restart isn't optional after an update: a browser refresh reloads the interface but not the node's HTTP routes, so a fresh panel can sit there doing nothing. New LoRA files are different - hit Rescan and the list updates live.

Where people get burned

Any top-level folder becomes a gallery, and previews are keyed to the file's path. A loras/utility/ folder shows up as a gallery named "utility" and loose files crowd into Unsorted; move or rename a LoRA later and its images are orphaned even though they're still on disk. The feature is only as good as your filing, so file things first.

Eight shots per LoRA per gallery, oldest dropped - including your cover, which falls back to the newest.

Trigger words are typed by hand. Nothing fetches them from CivitAI, so if you're used to that from LoRA Manager or rgthree this feels like a step backwards. The payoff: they come out as a wire you route, not a tooltip you retype.

The panel misbehaving after a ComfyUI update. Hard-refresh first (Ctrl/Cmd + Shift + R) - the front-end files cache aggressively. If the preview and buttons render wrong under the Nodes 2.0 beta, that's the known class of bug that hit rgthree's custom panels; turn the beta off and the legacy canvas still works.