Load unCLIP Checkpoint (Any)

This one's a bit of a time capsule. unCLIP checkpoints - the SD 2.1-unCLIP era, back around late 2022 - bundle a CLIP vision encoder directly into the checkpoint file, so the model itself can condition on a reference image rather than text alone. That made it the original "give me a variation of this photo" approach in the Stable Diffusion world, well before IP-Adapter and roughly two years before Flux Redux. unCLIPCheckpointLoader_Any is the loader for that specific checkpoint shape.

Why it has four outputs

Every other checkpoint loader in this pack gives you three outputs - model, CLIP, VAE. This one gives you four: MODEL, CLIP, VAE, and CLIP_VISION. That fourth output is the whole reason this node exists as separate from CheckpointLoaderSimple_Any - because unCLIP checkpoints carry their own CLIP vision encoder baked in, this is the only checkpoint loader in the pack where you don't need a separate CLIPVisionLoader_Any node alongside it. One file, one node, everything an image-conditioned unCLIP workflow needs.

The single required input is ckpt_name, picked from your models/checkpoints folder - same folder as any other checkpoint, but it needs to actually be an unCLIP-format file, not a regular SD checkpoint.

An honest gap

Worth saying plainly: this is the least-documented corner of the pack, and for good reason - unCLIP checkpoints are genuinely rare to encounter today. The identity-preservation and image-conditioning landscape has moved almost entirely to IP-Adapter, PuLID, and Flux Redux, all of which work as adapters layered onto a regular checkpoint rather than requiring a checkpoint trained specifically in the unCLIP format. If you're building a new image-variation workflow in 2026, you'll almost certainly reach for one of those instead. This node is here because the pack wraps every stock ComfyUI loader for consistency, not because unCLIP checkpoints are something you're likely to be hunting for.

The any input

Same mechanism as every other node in this pack: an optional any input, type accepts anything, and the value passed in is never used - it only exists to make this loader wait for whatever feeds it before running. ComfyUI would otherwise load this checkpoint (model, CLIP, VAE, and vision encoder all at once - a genuinely heavy bundle) as soon as nothing blocks it, typically immediately at graph start.

Installing it

ComfyUI Manager: search ComfyUI-LoaderUtils, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/lrzjason/ComfyUI-LoaderUtils

No extra dependencies - it wraps ComfyUI's own unCLIP checkpoint loader class.

Where people get burned

The most likely mistake is pointing this node at a regular SD checkpoint rather than an actual unCLIP-format file. Since this loader specifically expects the bundled vision-encoder architecture, feeding it a normal checkpoint either errors on load or leaves the CLIP_VISION output unusable - there's simply nothing there to extract. If you're not deliberately working with an unCLIP-format checkpoint, this almost certainly isn't the loader you want; CheckpointLoaderSimple_Any plus a separate CLIPVisionLoader_Any covers the modern equivalent of what this node does, and gives you more current model options on the vision-encoder side.