Icy CLIPTextEncode

IcyCLIPTextEncode is the IcyHider clone of the single most important node in ComfyUI: CLIPTextEncode. If you've ever stared at a downloaded workflow and wondered where the prompt lives - this is it. The author's own description says it best: "Encodes a text prompt using a CLIP model into an embedding that can be used to guide the diffusion model towards generating specific images."

What it does

Two inputs, one output. That's the whole node.

• text - "the text to be encoded," in a big multiline box. This is your prompt, as plain words.
• clip - "the CLIP model used for encoding the text," from your checkpoint loader or a CLIPLoader.

And the output - "a conditioning containing the embedded text used to guide the diffusion model" - is the CONDITIONING you wire into the sampler (and into anything downstream that modifies conditioning, like a ControlNet apply or a regional node).

Mechanically it's the translation step: the text encoder converts your words into a vector the diffusion model can steer by. Nothing downstream "reads" your prompt - the words never survive as text past this node. That's why changing the prompt text here is the only way to change what the model is guided by, and why the text is baked in at this exact point.

How to actually use it

Standard graph: two of these nodes sit between your checkpoint's CLIP and the sampler - one for the positive prompt ("a red fox in snow"), one for the negative ("blurry, low quality"). Both feed the sampler, and the model optimizes toward the positive and away from the negative.

A few things worth knowing:

• Token limits are real. CLIP's encoder caps out around 77 tokens; longer prompts get truncated. ComfyUI doesn't always warn you, so keep the important stuff early in the sentence.
• Weighting syntax exists but has an expiration date. (word:1.2)-style emphasis works on CLIP-encoded models. It does not work on the LLM text encoders (T5/Qwen) that modern models use - on those, the whole node may be replaced by a different encoder path.
• Negative prompts are inert on distilled models. On guidance-distilled checkpoints (most of what people run in 2026), the negative branch does nothing by default. Restate what you don't want as positive phrasing instead.

The "Icy" caveat

Outputs CONDITIONING, no image preview, so the Icy wrapper is pure relabeling - the pack clones every core node and this is one of the big ones that gets copied along. The encoding math is exactly core CLIPTextEncode; use whichever name keeps your graph readable.

Install

ComfyUI Manager → search "IcyHider" → install → restart, or:

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

Restart. No Python deps, no models bundled - the CLIP comes from whatever checkpoint/encoder you already load. Node under IcyHider Comfy Core. If your prompt seems to do nothing, check which branch this node is feeding and whether you're on a distilled model - those are the two classic culprits.