[SK] unCLIP Checkpoint

If you've never needed this node, that's normal - unCLIP checkpoints are a narrow, mostly legacy corner of the Stable Diffusion family (SD 2.1-unCLIP is the model most people have actually encountered), built for image-to-image variation that works off an image's visual embedding rather than a text prompt describing it. What tells them apart from an ordinary checkpoint is a fourth component baked into the file: a CLIP Vision encoder, alongside the usual MODEL, CLIP, and VAE. This node is the loader that knows to go looking for it. Point a regular SD1.5 or SDXL checkpoint at it and it simply won't have that fourth piece to hand back - this is specifically for files built with unCLIP in mind.

It's one of eight loaders in SK-ComfyUI-FolderingLoader, a small pack from GitHub author SKFRMSEHF whose whole reason for existing is a folder-tree file picker bolted onto rebuilds of ComfyUI's stock loaders. This is the unCLIP-flavored one - functionally the same idea as ComfyUI's own unCLIPCheckpointLoader, just with the tree browser instead of a flat dropdown. The pack ships no README and has no visible discussion anywhere online, so what follows leans on the node's own schema and general unCLIP mechanics rather than any pack-specific documentation, which simply doesn't exist yet.

How it works

Loading is the same split-the-archive-open operation as any checkpoint loader, just with one extra piece to extract. The CLIP Vision component is what makes unCLIP conditioning possible downstream: instead of (or alongside) a text prompt, you can feed it a reference image, and it encodes that image's visual concept into something the diffusion model can condition on - the mechanism behind "image variation" workflows where you want output that's visually in the same family as an input photo without describing it in words. That conditioning step happens in other nodes further down the graph (an unCLIP conditioning encoder); this node's whole job stops at handing you the four raw components.

The input and outputs that matter

There's one input - ckpt, picked from the folder-tree combo, listing whatever's in your checkpoints directory. It has no tooltip beyond "select file," so there's no extra parameter to tune here; the choice of file is the whole configuration.

Four outputs come out: MODEL and VAE wire exactly like any checkpoint's; CLIP goes to your text encode nodes if you're combining prompt and image conditioning; CLIP_VISION is the one that doesn't exist on a normal checkpoint loader, and it's what feeds an unCLIP conditioning node further downstream.

Installing it

No install section shipped with the pack, so it's the standard two paths: ComfyUI Manager → search SK-ComfyUI-FolderingLoader or the class name → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/SKFRMSEHF/comfyui_SK_Loader and restart manually. No bundled models - you supply the unCLIP checkpoint yourself, and those are a specific, less common download than an ordinary SD checkpoint, so double-check the model page actually says "unCLIP" before you grab it.

Common issues

• Wrong checkpoint type entirely. The single biggest trap: pointing this at a regular SD1.5/SDXL/Illustrious checkpoint. It'll load MODEL/CLIP/VAE fine and then have nothing sensible to put in CLIP_VISION, because that component genuinely isn't in the file. If you don't specifically need image-embedding conditioning, you almost certainly want SK_CheckpointLoaderSimple instead.
• Empty combo. Same as every loader in this pack - it only lists files already sitting in your checkpoints folder at startup, so add the file and restart rather than expecting a live refresh.
• This is a narrow, dated technique. unCLIP predates the newer instruction-editing and reference-image tooling that mostly replaced its use case; if you're trying to do "keep this character/composition, change the rest," look at what your model family's edit-conditioning tooling offers first - it's usually the more current answer.
• No community trail. Between this pack's obscurity and unCLIP's own niche status, there isn't a Reddit thread or wiki page waiting to bail you out. The node's own tooltips and the general unCLIP conditioning nodes in ComfyUI core are your best reference points.