ClipVision Select

CLIP-Vision is the encoder that turns an image into a conditioning vector - it's what makes IP-Adapter, image prompt nodes, and "reference this photo" workflows work at all. PgClipVisionSelect is the PG Nodes selector for it: a dropdown of the vision-encoder files in your install, outputting the chosen name so you can feed it into a CLIPVisionLoader or the Unified Loader's clip_vision slot.

This is a utility, and it stays a utility. Its whole job is to make the vision encoder's filename a first-class, visible choice in your graph instead of a value buried inside a loader. Useful when you're testing CLIP-Vision variants (ViT-L vs ViT-H vs the newer bigger encoders behave differently on the same reference image), or when you want one dropdown to drive both the loader and something else that needs to know which encoder is in play.

How it works

The list comes from ComfyUI's folder_paths registry, with the code trying several likely keys (clip_vision, clip_vision_models, vision_models, clip_vision_encoder) so it finds your files regardless of which folder convention your install uses. Selecting an entry returns that filename as the node's single output (clip_vision_name). As with all the PG selectors, the list is fixed at import time - drop a new encoder in and you'll need to restart ComfyUI before it appears.

Input: the clip_vision_name dropdown. Output: the chosen name, typed to the clip-vision list so it connects to a CLIPVisionLoader. The Uni variant (in PG/Select/UniLoader) adds a leading none so it can sit unused in a Unified Loader graph.

Install

Part of PG Nodes: ComfyUI Manager → "PG Nodes" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/GizmoR13/PG-Nodes

No models or dependencies.

Common issues

It's easy to confuse this with Clip Select - the former picks a text encoder, this one picks the vision encoder, and they feed different loaders. If the dropdown's empty, your CLIP-Vision files aren't in a folder ComfyUI scans (usually models/clip_vision); move them and restart. And remember the usual selector contract: this outputs a name, not the loaded model - the CLIPVisionLoader (or Unified Loader) downstream does the heavy lifting.