Anymatix CLIP Vision Loader

Text encoders turn your words into embeddings. A CLIP vision encoder turns a picture into embeddings - which is what lets a workflow say "make it look like this reference image" instead of describing everything in words. That's the whole job of the IP-Adapter family and of style/reference conditioning: encode the guide image with the vision model, then inject those features into the diffusion model alongside your prompt. AnymatixCLIPVisionLoader is the twin of ComfyUI's stock CLIPVisionLoader that loads that vision model, with the pack's usual twist - the model name is a string path rather than a dropdown.

It matters in the same URL-addressed way everything else in this pack works: a workflow that references a vision encoder by URL gets it downloaded by AnymatixFetcher and fed here as a path, so the graph opens and runs on machines that never saw the file before. The loader strips the input to a basename and lets ComfyUI resolve it against your models/clip_vision folder, so pasting a bare filename also works when the file's already local.

Where it sits in a workflow

The CLIP_VISION output doesn't feed a text encoder - it feeds a CLIPVisionEncode node, which turns the reference image into the conditioning your IP-Adapter or style-transfer nodes consume. If you've seen a Flux workflow with an "IPAdapter Style & Composition" block, the AnymatixCLIPVisionLoader is one of the quiet nodes feeding it. The vision encoder itself is a small-ish ViT-style model (typically a couple of GB), so it's cheap to keep loaded.

Inputs and output

• clip_name (STRING) - path or filename of the vision encoder weights (e.g. clip_vision_g.safetensors or a ViT-H one).
• Output: CLIP_VISION → CLIPVisionEncode.

One honest note for beginners: if you're only doing plain text-to-image, you don't need this node at all. It earns its place when your workflow conditions on an image - reference style, character consistency, IP-Adapter. If you can't see why a workflow would want it, you're probably looking at a workflow that does.

Install

Same pack story as its siblings:

cd ComfyUI/custom_nodes
git clone https://github.com/Anymatix/anymatix-comfy-nodes

then restart ComfyUI, or install through Manager by searching "anymatix-comfy-nodes".

Issues to expect

The classic stumble: loading a vision model that doesn't match what your IP-Adapter expects. The adapter weights and the vision encoder have to be the matching pair (a ViT-H adapter with a ViT-L encoder is a mismatch), so if your style transfer quietly produces mush, check the pairing before you blame the loader. And like every loader in this pack, it doesn't download anything itself - pair it with an AnymatixFetcher or have the file local, or you'll get a "file not found" that tells you exactly what happened.