Controlnet Models Selector v2

This is the settings box where the Searge SDXL workflow points at your actual ControlNet model files. It pairs with the SeargeControlnetAdapterV2 node: the adapter decides what kind of control to apply and to which image, this box decides which file on disk each mode uses. It writes those choices into the data-stream bus for the Magic Box to load.

It's the "Controlnet Models Selector v2" node, and it's pure plumbing - six dropdowns, one for each control type Searge supports, plus the CLIP-vision model that Revision needs. You set these once when you install the workflow and then forget they exist.

How it works

Searge's ControlNet support is built around Stability's official SDXL control-loras - canny, depth, recolor, and sketch - plus a custom slot and the CLIP-vision model for Revision. Each dropdown reads from your models/controlnet folder (or models/clip_vision for the vision model). When the adapter runs in, say, depth mode, the Magic Box grabs whatever you selected in the depth_checkpoint slot here.

The inputs and outputs that matter

Six required dropdowns, none of which do anything until you've downloaded the files:

• canny_checkpoint - the README recommends control-lora-canny-rank256.
• depth_checkpoint - control-lora-depth-rank256.
• recolor_checkpoint - control-lora-recolor-rank256.
• sketch_checkpoint - control-lora-sketch-rank256.
• clip_vision - clip_vision_g.safetensors, used by Revision rather than a spatial ControlNet.
• custom_checkpoint - a free slot for any other SDXL ControlNet you want to wire into the adapter's custom mode.

The one output is data (SRG_DATA_STREAM).

How to install it

Manager: search SeargeSDXL, install, restart. Manual, with the opencv dependency:

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

Restart ComfyUI, then download the models. The README links all four control-loras from Stability's control-lora HuggingFace repo (about 774 MB each) into models/controlnet, and clip_vision_g.safetensors (3.7 GB) into models/clip_vision. The adapter's preprocessors also want ControlNetHED.pth and res101.pth in models/annotators.

Common issues & troubleshooting

Dropdowns empty or files missing. This box only lists what's actually in your folders. If a slot is empty, you haven't downloaded that control-lora yet. None of these ship with the pack - they're separate multi-hundred-MB downloads, and the ControlNet features simply won't work until they're in place.

Name mismatch after import. Same story as every Searge selector: the workflow pre-fills specific file names, and if yours differ the slot falls back to the first available file. Open each dropdown and pick the right one.

The rank256 files are the affordable pick. Stability shipped these control-loras at multiple ranks; rank256 is the balance the README settled on and it's a sensible default. Bigger isn't dramatically better here, and given that SDXL ControlNets are the ecosystem's known weak spot regardless of file, there's little reason to hunt for exotic variants - depth and canny are the modes that actually hold up.