Icy CLIPVisionLoader

IcyCLIPVisionLoader is the IcyHider clone of CLIPVisionLoader, and it answers the question "where does the vision model come from?" in any workflow that uses reference images - IP-Adapter style transfer, style-model apply, that sort of thing. It's the quiet loading step behind the whole "make it look like this picture" family of tricks.

What it does

One input, one output:

• clip_name - the vision encoder file, from a dropdown populated by models/clip_vision. The folder matters: CLIP vision models live in their own directory, not models/checkpoints and not models/text_encoders. File in the wrong place = empty dropdown.

Output is a CLIP_VISION model, which you feed into CLIPVisionEncode (the node that actually reads an image and produces features). The loader and the encoder are a pair - this one brings the model into memory, the other one runs it over a picture.

The files you'll find here are things like clip-vit-large-patch14 (the SDXL-era vision encoder) and the newer vision encoders shipped for reference/IP-Adapter models. Sizes are modest compared to diffusion checkpoints, so loading one is cheap.

The typical chain

CLIPVisionLoader → CLIPVisionEncode → a transfer node (IP-Adapter, StyleModelApply) → sampler. That's the skeleton of every "match this reference image's look" workflow. The vision encoder's output features carry how the image looks, not what it literally is, which is exactly what the transfer node needs to inject style without copying content.

Where people get burned:

• Wrong folder. models/clip_vision specifically - the loader reads nothing else.
• Mismatched encoder and consumer. A vision encoder from an older generation feeding a newer transfer node can give weak results that look like "the feature isn't working." Match the version family if your workflow specifies one.
• No restart after dropping the file. ComfyUI scans model folders at startup; add a file while it's running and the dropdown stays empty until you restart.

The "Icy" caveat

Outputs CLIP_VISION - no image preview, so the Icy wrapper is pure relabeling. The pack clones every core node at startup and this loader gets copied along with the rest. The loading code is exactly core CLIPVisionLoader.

Install

ComfyUI Manager → search "IcyHider" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider.git

Restart. No Python deps, and no vision model ships with the pack - that file is yours to download and drop into models/clip_vision. Node under IcyHider Comfy Core. Empty dropdown, wrong folder, no restart: the eternal trio.