Nodepad_跳过CLIP文本层

Nodepad_跳过CLIP文本层 ("skip CLIP text layer") is the pack's rename of ComfyUI's core CLIPSetLastLayer node. Same mechanics, same inputs, Chinese label - and since this is one of the most misunderstood settings in the whole ecosystem, let's actually talk about what it does and when it matters, because the honest answer is "less than you'd think."

What it does

The CLIP text encoder (in SD 1.5 and SDXL, at least) is a stack of transformer layers. This node stops the encoder one or more layers early and conditions the image on that earlier, more general representation. The stop_at_clip_layer input is an integer from -24 to -1 (default -1), and the sign convention matters - this is ComfyUI's numbering. What AUTOMATIC1111 calls "CLIP skip 2" is -2 here. The author's tip field spells it out: -1 skips the last layer, -2 skips the second-to-last.

The inputs:

• clip - the CLIP from your checkpoint loader (or the pack's Nodepad_文本模型加载).
• stop_at_clip_layer - the skip value, negative, -1 default.
• tips - the author's explanation, decorative.

Output is a CLIP that you wire into your text-encode node, which then encodes your prompt against the truncated encoder.

Where it actually matters

The setting has exactly one strong use case left: SD 1.5 anime checkpoints and their lineage. The convention is 2 (that's -2 here) because NovelAI conditioned on CLIP's penultimate layer back in 2022 and the whole anime model family inherited that expectation. If you're running an Illustrious/Pony-family LoRA or checkpoint, -2 is usually the intended value, and a LoRA trained at skip 2 genuinely misbehaves if you run it at 1.

Everything else is where people waste time:

• SDXL models are already skipping. Both SDXL text encoders condition on the penultimate layer by default, so -2 restates the default and does nothing. Setting it to -1 is the one thing that actually breaks Pony/Illustrious - it turns them into noise.
• 2026's LLM-encoder models - Flux 2 Klein, Z-Image, Anima, Qwen-Image - have no CLIP at all. Their prompts are encoded by Qwen3 or Mistral, so there is literally nothing to skip; the input is inert. The community answer in 2026 is blunt: don't touch CLIP skip for any model except SD 1.5.

So the practical rule: set -2 for SD 1.5 anime work, and otherwise leave the default alone or don't use the node. If a modern workflow includes it, it's usually cargo cult from a copied template.

Installing the pack

• ComfyUI Manager: search "ComfyUI_AC_FUN" and install.
• Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/A719689614/ComfyUI_AC_FUN-V1.0

Restart ComfyUI. The pack is Cc啊程's (a Bilibili creator and architect by trade) Chinese-labeled wrapper suite; no model downloads, no heavy dependencies.

Bottom line

A faithful copy of a core node, useful precisely when core's CLIPSetLastLayer is useful (SD 1.5 anime at -2) and equally pointless everywhere else. If you're not running an anime SD 1.5 model, this node is decoration - but at least now you know it's supposed to be.