CLIP Selector

The core Load CLIP node does one job and does it fine: pick a text encoder file, get a CLIP object. CLIP Selector does that same job and then remembers which file you picked. In the kinorax pack, that difference is the whole point - the pack is built around carrying your model choices in image_info metadata, writing them into the saved image, and restoring them when you load the image back. A selector node is how your CLIP choice gets into that loop.

It's one of a family of selector nodes in the pack - Checkpoint, Diffusion Model, Unet, Lora, VAE, CLIP, plus Dual/Triple/Quadruple CLIP variants - all sharing the same design and all feeding the pack's loader pair (Load New Model / Use Loaded Model) or Image Info Context.

How it works

At execution time, CLIP Selector resolves the filename you chose in its dropdown against the actual text_encoders folder contents, packages it as an IPT-Clip value, and hands that to the loader pipeline. It's a choice-carrying node, not a loading node - the actual model gets loaded later, once, by Load New Model, so you don't pay for redundant loads when several workflow branches reference the same CLIP.

The two settings it passes through matter for the SD family:

• type - the CLIP type forwarded to the underlying loader (stable_diffusion is the default; SDXL/Flux use their own types via the dual variants).
• device - which device the CLIP runs on (default, or force CUDA/CPU).

Neither is something you'll touch much; they exist so the pack's loader calls the core loader correctly. The optional sha256 input is a cached hash used by the right-click View Model Info... action - the pack queries Civitai by SHA256 to show you metadata for the selected file, which is a nice touch for verifying you've got the exact model you think you do.

What you actually set

Just one field, honestly: clip_name, the dropdown of CLIP files in your text_encoders folder. If you keep CLIP-L or other text encoders there, they'll show up. If you're missing a file you expect, it means it's sitting in the wrong folder - the pack reads ComfyUI/models/text_encoders, not checkpoints.

Outputs are clip (the IPT-Clip value for the loader pipeline) and clip_name (a passthrough of the selected name, useful when you want the label downstream or in XY-plot axes).

Why bother versus core Load CLIP

If you never re-open your own old images, you can skip this node and you won't suffer. The payoff is the reproducibility story: load a saved image with Image Reader, and the CLIP choice comes back out of the metadata, ready to feed the loader again. For anyone doing batch runs across multiple text encoders, the selector also centralizes the choice - one dropdown at the top of the graph instead of hunting for loader nodes buried in the middle.

Installing

Part of kinorax/comfyui-info-prompt-toolkit. ComfyUI Manager, search "Info-Prompt-Toolkit":

cd ComfyUI/custom_nodes
git clone https://github.com/kinorax/comfyui-info-prompt-toolkit.git
cd comfyui-info-prompt-toolkit
pip install -r requirements.txt

Restart, and the selectors appear under Info-Prompt-Toolkit/ImageInfo. No extra model files - this node only references CLIPs you already have in text_encoders. If you're on an SDXL workflow with two encoders, grab Dual CLIP Selector instead; for a single-encoder Flux or SD1.5 setup, this is the one.