Checkpoint Loader (Searge)

This one's about as simple as ComfyUI nodes get: point it at a checkpoint file and it hands you back the three things every checkpoint contains - the MODEL, the CLIP, and the VAE. It's Searge's version of the loader you use at the start of essentially every workflow. If you've used ComfyUI's built-in Load Checkpoint (a.k.a. CheckpointLoaderSimple), this is the same node with a Searge label on it.

And that's the honest headline: it's in the pack's _deprecated_ category and does nothing the standard loader doesn't. It's kept around so old Searge workflows still open. For any new graph, use the built-in.

How it works

A .safetensors checkpoint bundles the diffusion UNet, the text encoder(s), and usually a VAE into one file. Loading it means unpacking those three components so the rest of your graph can wire to them separately - the model into a sampler, the CLIP into a text encoder, the VAE into a decode node. This node reads whatever's in your models/checkpoints folder and does exactly that unpack.

The inputs and outputs that matter

There's really only one thing to set:

• ckpt_name - the checkpoint file to load, picked from a dropdown of your models/checkpoints contents. For this pack that's typically an SDXL checkpoint like sd_xl_base_1.0_0.9vae.safetensors.

Three outputs:

• MODEL - the diffusion model, into your sampler (or a LoRA loader first).
• CLIP - the text encoder, into your prompt encoding.
• VAE - the decoder, into VAE Decode. Note this is the VAE embedded in the checkpoint, which for stock SDXL is the one that causes fp16 black images.

How to install it

Manager: search SeargeSDXL, install, restart. Manual, opencv dependency first:

python -m pip install opencv-python
cd ComfyUI/custom_nodes
git clone https://github.com/SeargeDP/SeargeSDXL.git

Restart ComfyUI. You'll want at least the SDXL base checkpoint (~7 GB) in models/checkpoints.

Common issues & troubleshooting

Black images from the embedded VAE. If you use this node's VAE output with stock SDXL, you can hit the classic fp16 black-image bug. The fix the README documents is to load a fixed external VAE (sdxl_vae.safetensors from the fp16-fix repo) separately and use that for decoding instead of the checkpoint's embedded one.

Dropdown empty. No checkpoints in models/checkpoints, or you didn't restart after adding one. Download a checkpoint and restart ComfyUI.

Prefer the built-in. Because this is deprecated, there's no reason to use it in a new workflow - ComfyUI's own Load Checkpoint is identical and current. And within the full Searge v4.x workflow, checkpoint loading is handled by the Model Selector box and the Magic Box's load checkpoints stage, not by this standalone node. This exists purely for backward compatibility with pre-v4 layouts.