Load CLIP Vision (Online)

CLIP vision is the "image in, embedding out" half of the CLIP family - the model that turns a picture into a conditioning vector so you can guide generation with an image rather than just text. Load CLIP Vision (Online) is the pack's wrapper for that encoder: pick it from the catalog, it downloads into your models/clip_vision folder, and you get a CLIP_VISION object out.

The catalog is even more minimal than the rest of the pack: exactly one entry, CLIP-ViT-bigG (the big OpenCLIP ViT-bigG/14 vision encoder from the LAION-2B training run). Which is convenient, because that's the one you want for most image-conditioning workflows anyway.

Where the output goes

One input, clip_name (the dropdown), and one output, CLIP_VISION. That output feeds into core ComfyUI nodes that take a vision encoder:

• CLIPVisionEncode - turns your reference image into CLIP_VISION_OUTPUT conditioning, which is the heart of IP-Adapter style transfer and image-guided workflows.
• unCLIP conditioning paths - if you're doing the image-prompted unCLIP style of generation, the vision encoder is the front door.
• Anything in the IPAdapter_plus ecosystem that wants a clip_vision socket.

Mechanically it's the pack's standard trick: on startup the pack pulls the catalog from api.aiart.doubtech.com/comfyui/clipvision, and the first time you pick CLIP-ViT-bigG it downloads the ~2 GB file into models/clip_vision, then hands the path to ComfyUI's stock clip_vision.load. No API key, no hosted inference - after the first fetch it's a local file like any other.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/yolanother/DTAIComfyLoaders

or search "DTAIComfyLoaders" ("Comfy UI Online Loaders") in ComfyUI Manager, restart. Only dependency is requests.

Where people get burned

• That first download is a big one. CLIP-ViT-bigG is roughly 2 GB, pulled over a Dropbox link with no progress bar in the UI. The node looks frozen while it works - it isn't.
• Dropbox HTML trap. Scripted downloads from Dropbox can come back as an HTML page instead of the model. If the load fails, check models/clip_vision: a tiny file means the download served HTML. Fetch the file manually and drop it in; the node will use it and stop re-downloading.
• Empty dropdown on fresh start. The catalog loads in a background thread; re-add the node after a moment, or restart.

Real talk: if you already have a CLIP vision encoder on disk for IP-Adapter (and if you run IP-Adapter you almost certainly do), the stock CLIPVisionLoader does exactly this. This node's only edge is skipping the download step on a fresh machine. For a first install it's a nice shortcut; for daily use it's a convenience you can take or leave.