Checkpoint Loader Ext

Let's get the uncomfortable truth out first: the CheckpointLoaderExt node itself is a renamed copy of ComfyUI's built-in CheckpointLoaderSimple. Same single input, same three outputs, same underlying load call. If you install this pack and just use the node, you've gained nothing. The "Ext" is the stuff bolted on around it - a tiny checkpoint manager with a web UI that remembers which of your models needs a separate VAE and what clip skip you like, then drops the right helper nodes next to your loader when you pick one. That part is genuinely handy.

Why you'd bother

You've got twenty checkpoints and you can't remember which ones have a baked-in VAE and which need a separate one, or that Counterfeit wants clip skip -2 and anything photoreal wants -1. The SD1.5/SDXL ecosystem is full of this friction - purple artifacts from a bad VAE, "why is my anime model washed out" (missing the clip skip), Illustrious-style models that loudly say don't attach a VAE on their CivitAI page. This node turns those per-model details into a little database instead of a browser tab full of model cards.

If you run one checkpoint, skip this pack. If you hoard fine-tunes, keep reading.

How it works

The loader half is dead simple. It reads ckpt_name from your models/checkpoints folder, calls load_checkpoint_guess_config with output_vae=True and output_clip=True, and hands you the first three results: MODEL, CLIP, VAE. Wire them like you would any checkpoint loader - MODEL to the sampler, CLIP to the positive/negative prompt encodes, VAE to the decoder. It behaves identically to the stock node, which is the point.

The actual trick lives in the bundled extension. On startup the pack creates a SQLite database (comfy_ui_checkpoints.db) and registers a handful of API routes on your ComfyUI server. Right-click the node → Checkpoints Info Manager opens the bundled web app at /art.quanse/web, where you catalog each checkpoint: version, download URL, whether it has a VAE baked in, clip skip, and a description. Then the frontend hooks the node's dropdown: when you change ckpt_name, it looks up your entry and acts. If the record says no VAE, it drops a VAELoader next to your node. If you set a clip skip, it writes that value into a CLIPSetLastLayer - updating one already wired to your CLIP output, or placing a fresh one beside the loader.

That clip-skip bit is the real win. ComfyUI's stock loader has no notion of "this model wants -2," so you always forget until your first garbage batch.

Installing it

ComfyUI Manager, search ComfyUI_CheckPointLoader_Ext, install, restart. Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/shuanshuan/ComfyUI_CheckPointLoader_Ext
# restart ComfyUI

There's no requirements.txt and nothing heavy - the pack uses aiohttp, sqlite3, and comfy.sd, all of which ComfyUI already ships. No model downloads, no API keys, no Python 3.12 drama. It even has a pyproject.toml for the Comfy Registry, which is why Manager finds it.

Where people get burned

• It's a stock loader until you fill in the catalog. Empty database means the node does nothing beyond CheckpointLoaderSimple. Populate the manager or you've installed a dependency for no reason.
• The database lives wherever you launch ComfyUI from. The code opens comfy_ui_checkpoints.db as a relative path, so it lands in your working directory - usually the ComfyUI install root. Launch from a different folder and you get a fresh, empty DB and your catalog looks gone. Back the file up with your config, not your models.
• The auto-dropped nodes aren't auto-wired (except the clip-skip value on a CLIPSetLastLayer already connected to your CLIP output). The VAELoader and fresh clip-skip node get placed beside the loader; you still connect the wires. It's a scaffold, not a miracle.
• The README is empty and the last commit is mid-2024. This is a personal-tool-of-a-pack by a single author (shuanshuan), no docs, and zero community chatter. For something this simple that's survivable - but don't expect fixes.

Honestly, it's a niche pack with a narrow job: remembering the VAE and clip-skip minutiae your workflow keeps tripping over. If that's your pain, it's worth the two-minute install.