PM Checkpoint Loader

If you've run a single ComfyUI workflow, you've met this node's boring cousin: the core "Load Checkpoint." It's the thing you drop at the top of the graph to get a model, a CLIP, and a VAE out of one file. PM Checkpoint Loader is the PM Manager pack's version of that - same job, friendlier face. Instead of a plain dropdown of filenames, you click a "select model" button and get a proper picker with previews and folder browsing, the kind of thing that stops mattering once your models are in one tidy folder and starts mattering the moment they aren't.

It's worth being clear about what it isn't. The description says "supports multiple models but only loads the selected one." The picker lets you browse everything you own, but the node loads exactly the checkpoint you've highlighted, one at a time. This is not a batch loader and not a model switcher - if you want to flip between models, you still wire up a switch.

How it works

Behind the scenes it's the same machinery as core ComfyUI. The picker UI feeds a hidden widget, and the node calls comfy.sd.load_checkpoint_guess_config on the file you chose, which auto-detects the architecture and hands back a model, a clip, and a vae. If the node can't find the selected file, it raises a clear "No checkpoint model selected" / "not found" error instead of failing silently.

The one knob you actually set is weight_dtype, a dropdown with four options: default, fp8_e4m3fn, fp8_e4m3fn_fast, and fp8_e5m2. Those are the two fp8 flavors. e4m3fn is the usual quantized format - the one people mean when they say "fp8" - and _fast adds the fp8 optimization passes that make it speed up as well as slim down. e5m2 trades a bit of precision for more exponent range; you almost never need it. If your card is tight on VRAM, fp8_e4m3fn is the safe first stop; if it fits, default is fine and you don't have to think about it.

Inputs and outputs

• weight_dtype - the only input, and it's optional in spirit (default = no quantization).
• model - wire this into your KSampler.
• clip - wire this into a CLIP Text Encode for your positive and negative prompts.
• vae - wire this into VAE Decode to turn the latent back into pixels.
• model_name - a string with the checkpoint's filename. It's a * type, so it won't fight you when you wire it into filename or metadata nodes.

Installing it

This is a PM Manager pack node, so the install is the pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/pixel-magician/ComfyUI-PM-Manager.git
# then restart ComfyUI

Or, easier, use ComfyUI Manager: search for ComfyUI-PM-Manager in the Manager's install tab and hit install. Its requirements (opencv-python and imageio-ffmpeg) come along automatically. No model files to download - it only manages the checkpoints you already own.

Common issues

The most reliable way to get bitten by this pack is version. It's built on ComfyUI's newer extension API (comfy_api.latest), so if the PM nodes don't show up in your node list at all, the answer is almost always "update ComfyUI," not "reinstall the pack." If you get a "No checkpoint model selected" error, the picker widget didn't populate - refresh the browser tab and re-pick. And remember it loads one checkpoint: if you expected the other file to also be in memory, it isn't.