Icy unCLIPCheckpointLoader

Most checkpoints give you three outputs: MODEL, CLIP, and VAE. unCLIP checkpoints give you a fourth - CLIP_VISION - and that extra output is the whole reason this loader exists. It's the node for a small, specific family of SD 2.x-era models (the "unCLIP" variants, like SD 2.1-unclip) that can take a reference image as conditioning in addition to text. Think of it as image-prompting with a checkpoint that was built for it, before IP-Adapter made that trick mainstream.

The "Icy" prefix is the pack's privacy coat, not a different loader. IcyHider subclasses core ComfyUI nodes at startup - same inputs, outputs, and behavior, filed under the IcyHider Comfy Core category - and its JavaScript hides previews until hover. This is core unCLIPCheckpointLoader with a cover on.

Input and outputs

From the schema, one input and four outputs:

• ckpt_name - dropdown of checkpoint files in ComfyUI/models/checkpoints. The unCLIP models land here like any checkpoint.
• MODEL, CLIP, VAE - the standard trio, wired exactly like a normal CheckpointLoader.
• CLIP_VISION - the extra one. This is the vision encoder that reads reference images, and it feeds into the unCLIPConditioning node.

When you'd actually use this

The graph is unusual: instead of LoadImage → VAE Encode → img2img, you feed a reference image through a CLIP Vision encoder, and that visual description becomes part of the conditioning alongside your text. The result is a form of image-guided generation that works at full denoise - the image informs the composition without the pixel-locking behavior of img2img.

Honest framing: this whole path is a museum piece in 2026. The unCLIP checkpoints are SD 2.x-era, which means they're an older model family that most of the current ecosystem has moved past, and IP-Adapter plus reference conditioning covers the same idea on modern models with better results. If you're here it's because a workflow you found uses it, or you're curious about the history. Either way: it loads like any checkpoint, and the CLIP_VISION output is what makes it special. If a workflow asks for this node and you've swapped in a plain SD 2.1 checkpoint, the CLIP_VISION wire will be missing - that's the tell that you grabbed the wrong file.

Install

Frontend-only pack, no dependencies:

# ComfyUI Manager: search "IcyHider", Install, restart

# or:
cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider.git
# restart ComfyUI

Pack-wide: the hiding is DOM-based, so the newer Nodes 2.0 canvas frontend may not engage it, and the cover is screen-only - the checkpoint file on disk is real. If the ckpt_name dropdown is empty, drop an unCLIP checkpoint into ComfyUI/models/checkpoints and refresh.