Icy unCLIPConditioning

IcyunCLIPConditioning is the node that lets a reference image become part of your prompt. It takes your text conditioning, pairs it with the output of a CLIP vision encoder (which has "seen" a reference image), and produces combined conditioning the sampler can use. The upshot: you can generate something that follows a reference's vibe or composition at full denoise, rather than smearing it through an img2img pass.

The "Icy" prefix is the pack's privacy coat, not a different node. 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 unCLIPConditioning with a cover on. And honestly, in a privacy pack, this is one of the more useful wrappers: the reference image that drives this node is frequently the thing you don't want on screen.

The inputs that matter

From the schema, four:

• conditioning - your text conditioning from a CLIPTextEncode.
• clip_vision_output - the CLIP_VISION_OUTPUT from a CLIPVisionLoader or from an unCLIP checkpoint's vision path. This is where the reference image enters.
• strength - how hard the image conditioning pushes, default 1.0. Drop it to 0.5 for a subtler influence.
• noise_augmentation - how much noise is added to the image conditioning, default 0. This is the interesting dial: higher values loosen the image's grip on exact layout while keeping its character. Crank it and the result is more "inspired by" than "following."

One output: CONDITIONING, ready for a sampler.

How it fits together

The unCLIP path is a small, self-contained subgraph: an unCLIP checkpoint loader (or a CLIPVisionLoader) produces the vision output; this node merges it with text conditioning; the sampler does the rest. It was the original "image-prompting" mechanism in the SD 2.x era - the CLIP vision encoder converts the reference into embeddings that the unCLIP models were trained to follow.

The noise_augmentation knob deserves a sentence because it's the one people skip: at 0, the reference often dominates composition tightly; as you raise it, you trade fidelity for variety. It's the closest thing this old mechanism has to a creativity slider. And a quick honest note - this whole family is legacy in 2026, superseded by IP-Adapter and modern reference conditioning. If a workflow you downloaded uses it, it works fine, but don't go building new pipelines on it.

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 cover is DOM-based, so the newer Nodes 2.0 canvas frontend may not engage it, and it's screen-only privacy. If your sampler output ignores the reference entirely, the usual culprit is a missing or miswired clip_vision_output - that wire is the whole feature.