CCSR_Model_Select

This is the older of the two model loaders in Kijai's ComfyUI-CCSR pack, and the README is upfront about what it is: the original node, kept around "for backwards compatibility" - Kijai won't change it. CCSR_Model_Select opens a dropdown of everything in your ComfyUI/models/checkpoints folder, loads the CCSR checkpoint you point it at, and outputs a ccsr_model that feeds CCSR_Upscale.

It exists because when the pack launched, CCSR only shipped as a full-precision checkpoint from the project's Google Drive, and that's where the README told you to put it. The modern path is DownloadAndLoadCCSRModel, which pulls the fp16 safetensors from HuggingFace automatically and is strictly easier. You'd reach for this one in two situations: you already have the original checkpoint sitting in checkpoints, or you're following an old workflow that calls it by name.

How it works

Load-wise it's the classic pattern: build the model from the pack's stage-2 config, then strict-load the checkpoint's state dict. One behavior worth knowing before you trip on it: the model is cached on the node instance. It loads once per session and stays in memory, which is why the pack got the "model loading is twice as fast" note in the README - but it also means switching ckpt_name mid-session does nothing. You pick a model at the start of a session and it sticks until you restart ComfyUI. There's no way to force a reload from the UI, so factor that into how you use it.

It defaults to fp16 when your setup supports it, same as the rest of the pack.

The one real trap

The dropdown lists every file in models/checkpoints - all your SD checkpoints, LoRAs-as-checkpoints, everything. Grab the wrong one and you get a load error, because strict=True rejects a state dict that doesn't match CCSR's architecture. It's not a broken pack; it's a dropdown that can't tell a checkpoint from a model. The download-and-load node doesn't have this problem because it only ever shows the two correct files.

Installing

Same pack install as everything else in ComfyUI-CCSR - ComfyUI Manager (search "ComfyUI-CCSR") or:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-CCSR

Restart and let Manager install the requirements: taming-transformers, omegaconf, einops, pytorch-lightning>=2.2.1, open-clip-torch>=2.23.0.

Should you use it at all?

Honestly, probably not, unless you're stuck with the old checkpoint. DownloadAndLoadCCSRModel is the same output with zero manual file management. But if a 2024 tutorial or an old shared workflow drops this node on your graph, now you know it isn't broken - it's just the pack's history wearing a dropdown.