Neons Model Explorer

If your checkpoints folder has thirty files and half of them are named after an anime character you downloaded in a haze nine months ago, you know the problem. ComfyUI's stock loader gives you a dropdown of filenames. Neons Model Explorer gives you the same dropdown with a picture next to each entry and a note field that comes back out as a string - the sampler, the CFG, the resolution that checkpoint actually wants.

It's one of five nodes in the Neons Style Explorer pack - really a 3,015-entry style-prefix catalog that happens to ship loaders - and it's the one you'd reach for on a day when you never touch a style.

What it does differently

Mechanically it replaces Load Checkpoint and UNETLoader with one node that reads both checkpoints and diffusion_models. Names carry their source, so diffusion_models/flux/flux1-dev.safetensors and a checkpoint of the same name are separate entries with separate previews, and the browser's dropdown switches between the two - your SDXL pile and your Flux UNETs never mix.

The filing is free: the folders you already have are the galleries. A top-level folder is a gallery, a folder inside it a family, loose files land in Unsorted. Previews follow the file rather than the path - each is recorded against a fingerprint of the file, not its location - so moving a checkpoint keeps its images, while a different file dropped into the old spot doesn't inherit them. Storage sits under user/models/<gallery>/, with up to eight shots per checkpoint and a cover you pick by clicking a thumbnail. The gallery ships empty. You fill it.

The notes are the part I'd keep. A checkpoint doesn't want extra words in your prompt the way a LoRA does, but it usually wants a reminder, and sampler/CFG advice isn't transferable - a distilled 8-step model and a CFG 4–7 SDXL finetune want opposite things. Type dpmpp_2m, cfg 4.5, 1024 under the preview and wire the notes output wherever you like. Same move rgthree's Power LoRA Loader made with trigger words, minus the prompt pollution.

The inputs and outputs that matter

One input, and you can't get it wrong:

• ckpt_name - the picker, built at runtime from your checkpoints and diffusion_models folders with None at the top. Its own tooltip says it best: a diffusion model has no CLIP or VAE of its own, so those outputs stay empty and you wire your own loaders.

Five outputs:

• model - into your KSampler. Always populated.
• clip, vae - into the text encode and VAE Decode. Filled for a checkpoint by ComfyUI's usual guess-config path, embeddings included; None for anything out of diffusion_models.
• model_name - a plain string, handy for filename prefixes and logging which model made which image.
• notes - the text you typed under the preview.

ckpt_name is a widget, so right-click → Convert widget to input lets you drive it from a string. That's the whole mechanic for swapping models across a batch without hand-editing the graph.

Install

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

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

Restart, then find it under right-click → Add Node → Neons. No extra dependencies - Pillow and numpy, both already in ComfyUI. Python 3.9+, MIT licensed.

Where people get burned

Restart, don't refresh. This is the pack's single biggest source of confusion. A browser refresh loads the new front end, but the node's HTTP routes only register when the ComfyUI process starts, so you get a new-looking panel talking to an old server and anything new quietly fails. If a control tells you the running server hasn't loaded this version, that's what happened.

Empty clip/vae is not a bug. You picked something out of diffusion_models, and the file is the UNET alone. Wire a CLIP loader and a VAE loader, and set the CLIP type yourself.

A stale or moved entry. The picker comes from ComfyUI's folder scan, so a checkpoint you dropped in while ComfyUI was running may need a node-definition refresh. Queue with None selected and you get a flat "choose a model" error; move a file afterwards and the node says so rather than guessing.

Stale notes. Editing a note leaves every widget exactly as it was, and ComfyUI caches nodes whose inputs haven't changed - so the notes output used to hand out the old text. The cache signature now folds the note in: an edit re-runs the node, and choosing the same model twice still caches.

A preview that won't save. Save is manual here, with no crawl and no auto-populate, because swapping checkpoints mid-batch reloads gigabytes. When it can't identify an image from that prompt it says so on the preview instead of filing it against the wrong model.

The pack has almost no Reddit footprint - I went looking and found nothing under its own name, which is normal for a young single-author pack. So you're not inheriting folklore here, just a loader with pictures and a sticky note.