Nodes/ComfyUI-IcyHider/Icy CLIPSetLastLayer
ComfyUI Node

Icy CLIPSetLastLayer

CLIP skip, in node form — the -2 dial that anime models expect

By icekiub-ai·Created 9 months ago·Updated 9 months ago· 31
Icy CLIPSetLastLayer
  • clip
  • CLIP
stop_at_clip_layer-1

IcyCLIPSetLastLayer is the IcyHider clone of CLIPSetLastLayer - the node form of what the A1111 world calls "CLIP skip." It's a small node with an outsized reputation, and it's also a setting whose ground has been shifting under everyone's feet.

What it does

Two inputs:

  • clip - the CLIP text encoder (from a checkpoint or a CLIPLoader).
  • stop_at_clip_layer - an integer from -24 to -1, default -1. Negative indexing is the whole UI: -1 means "use all layers," -2 means "stop one layer early," and so on. It re-emits the same CLIP with a different last layer.

Output is a CLIP that feeds your CLIPTextEncode. That's the node. Everything interesting is in what "stopping a layer early" does.

Why it exists

CLIP's text encoder (ViT-L/14 in SD 1.5) has 12 layers, and later layers hold more specific, more literal detail while earlier layers hold more general categories. "Stop a layer early" = condition on the more general representation. That was a deliberate design choice: NovelAI conditioned on the penultimate layer back in 2022, its weights became the ancestor of most anime checkpoints, and every descendant through Pony, Illustrious, and NoobAI inherited the expectation. So on an anime SD 1.5 model, the standing recipe is -2 - that's what "Clip Skip: 2" on release posts means - while realistic SD 1.5 models usually run the default -1.

Here's where the folklore breaks, and it's worth knowing:

  • SDXL already skips. ComfyUI initializes SDXL's CLIP-L at layer -2 and its CLIP-G at the penultimate layer by default. Setting -2 on SDXL often changes nothing; setting -1 on a Pony-family model is what turns it into noise. The advice "set CLIP skip 2" on SDXL is really "don't set it to 1."
  • Modern models have no CLIP to skip. Flux, Z-Image, Wan, and friends encode with LLM text encoders (T5, Qwen, Mistral). There's no CLIP, no layer ladder, and this setting does not exist for them. If you're on a 2026 base model, this node is a no-op at best.

So the practical rule: SD 1.5 anime model → -2. SD 1.5 realistic → -1. SDXL → leave it. Anything with an LLM encoder → don't touch it.

The "Icy" caveat

Outputs CLIP - no preview for IcyHider to hide, so the Icy prefix is cosmetic relabeling of the core node. Same math, same behavior.

Install

ComfyUI Manager → "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. Node under IcyHider Comfy Core. And if you're debugging why a Pony workflow looks wrong, check this node's value first - it's the quiet culprit more often than people expect.

CategoryIcyHider Comfy Core

Inputs (2)

NameTypeDefaultDescription
clipCLIP
stop_at_clip_layerINT-1-24–-1

Outputs (1)

NameTypeDescription
CLIPCLIP