IPA_clip_vision

Before you can hand a reference image to an IP-Adapter-style node and say "make it look like this," something has to turn that picture into an embedding the model can actually use. IPA_clip_vision is that first step - it's this pack's version of a CLIP vision encoder, living in its chx_IPA tool group alongside the pack's other image-prompting nodes (the README lists "Ipa style" and "redux transfer" among the things its general control stack handles).

How it works

IP-Adapter's whole trick is that it doesn't fine-tune anything - it's a small (~22M parameter) adapter that injects a CLIP image embedding into the model's cross-attention alongside the normal text conditioning, so your text prompt keeps working and the image just adds on top of it. But that only works if the reference image has already been run through a matching CLIP vision model to produce the embedding in the first place. That's the entire job of this node: pick a CLIP vision model, feed it an image, get back a CLIP_VISION_OUTPUT object that a downstream IP-Adapter-apply node knows how to consume.

Inputs and outputs

Two inputs, both required:

• clip_name - a dropdown of CLIP vision models found in your models/clip_vision folder. If it comes up empty, that's not a bug in the node - it means you haven't downloaded a CLIP vision checkpoint yet (see below).
• image - the reference picture to encode.

One output: CLIP_VISION_OUTPUT, which only makes sense wired into another IP-Adapter-family node downstream that expects that exact type - it's not an image, a latent, or anything you'd preview directly.

How to install it

ComfyUI Manager → search ComfyUI-Apt_Preset, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git

then run install.bat for dependencies (Windows-only - on Linux/macOS run its pip install lines by hand) and restart ComfyUI.

You also need an actual CLIP vision model file, which this node doesn't download for you. The standard ones people use for IP-Adapter work are OpenCLIP ViT-H-14 for SD 1.5 checkpoints and ViT-bigG-14 for SDXL - grab the matching one from Hugging Face and drop it in ComfyUI/models/clip_vision/.

Common issues & troubleshooting

clip_name dropdown is empty. This means ComfyUI can't find any files in models/clip_vision - the node isn't broken, there's just nothing there yet. Download a CLIP vision model and restart, or refresh the node list.

Using the wrong CLIP vision model for your base checkpoint. SD 1.5 and SDXL expect different CLIP vision encoders (ViT-H-14 vs. ViT-bigG-14) - mismatching them typically produces garbled or ineffective image conditioning rather than a hard error, which makes it a quiet failure mode worth double-checking if IP-Adapter-style results look wrong.

Worth knowing going in: IP-Adapter itself is a mature, largely-frozen technology. Upstream development on the original project stopped in January 2024, and the community node pack most people associate with it has been in maintenance mode for a while. It still works fine on SD 1.5 and SDXL - it just isn't where new development attention goes; newer architectures like Flux route equivalent image-prompting jobs through different tools (Redux, PuLID, in-context editors) rather than IP-Adapter.

IMPORT FAILED on startup. This pack has at least one reported import failure in the wild on a stock install, generally traced to skipping install.bat's dependency step - which has to be run by hand off Windows. Check the console for the actual missing package before reinstalling.