MinusZone - KolorsCLIPVisionLoader - Legacy

MZ_KolorsCLIPVisionLoader loads the vision model that Kolors' IP-Adapter uses to read your reference image. In the legacy Kolors IP-Adapter graph it's the node that produces the CLIP_VISION object - the piece that turns a photo into the tokens the adapter injects. It's flagged "Legacy" for a good reason: once the pack moved IP-Adapter support over to cubiq's ComfyUI_IPAdapter_plus, the clip vision loading moved with it, and this node became a shortcut for old workflows that reference it by name.

What it loads

The input, clip_name, is a dropdown of everything in models/clip_vision/. For Kolors IP-Adapter you want the image encoder that shipped with the official adapter - the pytorch_model.bin from Kwai-Kolors/Kolors-IP-Adapter-Plus's image_encoder/ folder, which is a CLIP ViT variant (the clip-vit-large-patch14-336 build used by the FaceID line). Drop that in models/clip_vision/, pick it, and the node loads it via ComfyUI's standard clip vision loader, inside the pack's Kolors hook context. Output: a single CLIP_VISION.

How it fits

In the legacy chain it feeds the clip_vision input of MZ_IPAdapterAdvancedKolors (or the FaceID node). In practice, the IPADAPTER object from MZ_IPAdapterModelLoaderKolors already carries the clip vision reference, so this node is the explicit way to load or override it when you're building the graph by hand. If you're assembling a Kolors IP-Adapter workflow from scratch today, don't bother with this node - cubiq's ComfyUI_IPAdapter_plus handles clip vision internally. Keep it for reviving saved workflows that reference "KolorsCLIPVisionLoader".

Install

Standard pack install: ComfyUI Manager search "ComfyUI-Kolors-MZ", or git clone https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ into custom_nodes and restart. Then:

# from Kwai-Kolors/Kolors-IP-Adapter-Plus
image_encoder/pytorch_model.bin -> models/clip_vision/

Common issues

The classic failure is an empty dropdown or a load error because the file landed in the wrong place - models/clip_vision/ is the only folder this node reads. Wrong-architecture clip vision files (e.g., a regular SDXL ViT-bigG build) will usually still load but produce weak or wrong image conditioning; you want the encoder that came with the Kolors adapter. And keep the version note in mind: this is a Legacy node, so if ComfyUI updates break it, the supported path is cubiq's pack, not a fix here.