Model Selector v2

This is the box in the Searge SDXL workflow where you tell it which checkpoints to run. Base model, refiner, VAE - three dropdowns, and it hands the choices off to the rest of the pipeline through Searge's data-stream bus. It's the "Model Selector v2" node, and it's one of the first things you'll touch when you set up the workflow, because the defaults point at file names you probably don't have yet.

Worth knowing up front: this isn't a general-purpose checkpoint loader you'd drop into any graph. It only makes sense inside the Searge workflow, because its only output is an SRG_DATA_STREAM - Searge's private "settings bus" - not a MODEL you can wire into a KSampler. If you want a plain loader, ComfyUI's built-in Load Checkpoint is what you want. This node exists to feed the Magic Box engine.

How it works

SDXL was designed as a two-model pipeline: a base checkpoint that does the heavy lifting, and a refiner checkpoint meant to add fine detail in a short second pass. Searge's workflow was built around that two-stage design when it was still the intended way to run SDXL. So this selector gives you a slot for each, plus a VAE choice. Whatever you pick gets written into the data stream and the load checkpoints stage of the Magic Box reads it back out and actually loads the files.

The inputs and outputs that matter

Three required dropdowns:

• base_checkpoint - your SDXL base model. The README pre-selects sd_xl_base_1.0_0.9vae.safetensors, so if you downloaded that (7 GB, from Stability's HuggingFace) it should just be there.
• refiner_checkpoint - the SDXL refiner, or none if you'd rather skip it. Given that the community quietly dropped the refiner as barely worth the extra step, none is a defensible choice.
• vae_checkpoint - either from base model, from refiner model, or an external VAE file. This one matters more than it looks: SDXL's embedded VAE is the one that produces black images in fp16, and picking a fixed external VAE here is the fix.

The single output is data (SRG_DATA_STREAM), which chains into the next settings box or into the Magic Box.

How to install it

ComfyUI Manager: search SeargeSDXL, install, restart. Manually, and don't skip the opencv dependency the README requires:

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

Restart ComfyUI. Then grab at least the SDXL base checkpoint into models/checkpoints; the refiner and the fixed VAE are optional-but-recommended per the README.

Common issues & troubleshooting

The dropdowns are empty or show the wrong file. The workflow ships with specific file names pre-filled. If your files are named differently, the selector falls back to whatever's first in your list - click each dropdown and pick your actual checkpoint. The README explicitly warns you'll need to fix these names by hand.

Black images out of the workflow. Classic SDXL fp16 VAE bug. Download the fixed sdxl_vae.safetensors (the madebyollin/sdxl-vae-fp16-fix one the README links), drop it in models/vae, and select it here instead of the embedded VAE. This is the documented fix.

Refiner slot confusion. You don't have to use a refiner. Set refiner_checkpoint to none and the workflow runs base-only, which is how most people run SDXL in 2026 anyway. The refiner was a 2023 idea that didn't stick.