DTunCLIPCheckpointLoader

unCLIP is the image-prompted cousin of Stable Diffusion - the SD 2.1-era family where, instead of only typing a prompt, you feed the model a reference image and it generates "in the vibe of" that image. The checkpoints that ship this capability are special: alongside the usual MODEL, CLIP, and VAE, they include a CLIP vision encoder baked in, which is how the image gets turned into conditioning. That's the whole reason this loader has four outputs where every other checkpoint loader in the pack has three.

DTunCLIPCheckpointLoader is the pack's wrapper for that family. Pick a checkpoint from the catalog, it downloads into models/checkpoints, and out come MODEL, CLIP, VAE, and - the differentiator - CLIP_VISION.

The catalog

The dropdown (ckpt_name) lists six entries, all from comfyanonymous's wd-1.5-beta2_unCLIP HuggingFace repo (direct links, so these downloads are the reliable kind in this pack): Asthetic Unclip H/L and Beta 2 Unclip, each in FP16 and FP32. Two notes:

"Asthetic" is the author's spelling, not yours - that's the "aesthetic" fine-tune, tuned for image-guided aesthetic transfer.
FP16 is the pragmatic pick unless you have a precision reason not to; roughly half the download and no meaningful quality difference at inference, per the KB's fp16 notes.

How to wire the four outputs

MODEL → the KSampler, as usual.
CLIP → CLIP Text Encode for your prompt.
VAE → VAE Encode, where you feed the reference image that will drive the generation.
CLIP_VISION → CLIPVisionEncode, which turns that same reference image into image conditioning for the sampler.

The image does double duty: VAE Encode puts its pixels into the latent, and the vision encoder captures its "meaning" as conditioning. That pairing is the unCLIP mechanism - the model reinterprets rather than copies.

How it works

Standard pack machinery (catalog from api.aiart.doubtech.com/comfyui/unclip, on-demand download), loading through ComfyUI's guess-config checkpoint path with output_clipvision=True. The one wrinkle is disk: these are full checkpoints, so the first download is a GB-scale wait with no progress bar.

Install

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

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

Where people get burned

Forgetting the fourth output. If you just wire MODEL/CLIP/VAE and skip CLIP_VISION, you've built a normal text-to-image workflow and the unCLIP checkpoint does nothing special. The vision encoder is the point.
Pairing with a wrong VAE. The VAE output here matches this checkpoint family - don't swap in an SDXL VAE or the decode falls apart.
Legacy expectations. unCLIP is a 2022–23 technique; the results are more "stylistically influenced" than photorealistic transfer. The KB's forgotten-models essay treats this family as part of the era's interesting-but-superseded work. If your goal is modern image-to-image, IP-Adapter or a controlnet pipeline is the current tool.

This is the most interesting node in the pack - four outputs, a genuinely different loading pattern, and a catalog that actually has content. It's a great way to poke at image-prompted generation without hunting down legacy checkpoints yourself. Just remember to wire that fourth socket.