🧬 Checkpoint Discovery Hub

If you juggle more than a few models, you know the drill: every time you want to try a different checkpoint you open the UNET loader dropdown, scroll a long list of nearly identical filenames, and hope you picked the right one. 🧬 Checkpoint Discovery Hub is a bet that this is a solved problem waiting for a better UI. It replaces the usual loader stack - UNET loader, dual CLIP loader, VAE loader - with one node whose front end is a full model gallery: thumbnails, folder filters, favorites, presets, and a Civitai sync button. Pick a card, and the node hands you a ready-to-wire MODEL, CLIP, and VAE.

The honest framing first: this is a young, small pack. Roughly three dozen Google impressions on comfy.icu and no real community buzz behind it yet, so treat it as a quality-of-life experiment rather than a community staple like rgthree or Efficiency Nodes. What it does well, though, is close the gap between "I know what model I want" and "I can see it before I load it."

How it works

Mechanically it's not magic - that's the reassuring part. The node has no input sockets at all; everything you configure lives in the node's own UI, and the UI's state is passed into the node as a JSON blob. Inside, it calls the same ComfyUI core loaders you'd wire by hand: the UNET gets loaded from your models/diffusion_models folder (with optional fp8 casting), one or two CLIPs from models/text_encoders, and a VAE from models/vae. The gallery also offers the TAESD "approximate" VAEs (taesd, taesdxl, taesd3, taef1) plus a pixel_space passthrough - handy when you just want a fast preview and don't care about perfect decode quality.

Because the node's IS_CHANGED key is the current selection, switching models forces the downstream graph to actually re-run instead of being skipped as cached. Your selections, favorites, presets, and Civitai metadata all live in small JSON files inside the pack's folder, and the UI state restores after a restart.

The Civitai sync is the one genuinely clever bit: it SHA-256 hashes your local model, asks Civitai's by-hash API which version that is, then downloads the preview image (transcoded to width 450) and saves it next to the model file so your gallery shows thumbnails. Needs internet, and it writes into your model folders - know that before you click it.

The outputs that matter

All four outputs are there for the taking:

• MODÈLE - the MODEL tensor, wire it into a sampler. (Yes, the output is called "MODÈLE" - the author is French, the code comments are French, and the accent traveled. It's the MODEL you're used to.)
• CLIP - your text encoder(s), wired to a conditioning node.
• VAE - decode your latents at the end of the graph.
• model_name - the checkpoint's filename as a string. Genuinely useful: feed it to a save node, a display node, or any custom node that wants to know which model produced an image.

That model_name string is the sleeper feature - most loaders make you re-type or hardcode the name elsewhere, and here it just follows whatever you clicked.

Install

Two ways, both trivial:

# ComfyUI Manager: search "ComfyUI-checkpoint-Discovery-Hub" and hit install.

# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/Light-x02/ComfyUI-checkpoint-Discovery-Hub.git

Restart ComfyUI after either. There are no heavy dependencies - the requirements (numpy, Pillow, aiohttp, requests, tqdm) are things ComfyUI already pulls in, and there are no model files to download. This is one of the cleaner installs in the ecosystem.

Gotchas

The big one: the gallery only browses models/diffusion_models. If you're still running classic combined checkpoints from models/checkpoints, they won't appear - drop them into diffusion_models (or add the folder in extra_model_paths.yaml) and they'll show up. Also, the presets and favorites JSON files live inside the pack's directory, so a git clean or deleting the folder wipes them - back them up if you build a big preset library. And remember the sync behavior above: hitting that Civitai button puts preview files and metadata in your model folders.

Bottom line

If you have three models, this is overkill - the built-in loader is fine. If you have thirty, or you switch between Flux fp8 and SDXL setups constantly, the gallery, presets, and favorites earn their keep, and model_name alone justifies the install. It's a one-author pack with no track record yet, so watch for updates - but the mechanism is thin enough that it can't easily break your other workflows.