PROMPT GUIDE

What Is Clip Skip? Its Meaning and How to Choose 1 or 2

Stable Diffusion settings include an item called "Clip skip." Many people heard that "setting it to 2 makes it prettier," changed the number, and could not tell what changed. This guide sums up what Clip Skip does, how to decide between 1 and 2, and how to review things when a picture looks wrong.

Last updated: 2026-10-22 approx. 3 min read

日本語版

The short answer

Clip Skip decides which layer, counting from the back, of the prompt-reading part (CLIP) is used. 1 uses it to the end, and 2 stops one layer earlier. The basic rule is to match how the model was trained, so use the value written on the model page, and if none is written, line up 1 and 2 with the same seed and compare.

What is Clip Skip?

It is a setting that decides how many of the last layers of CLIP, which reads the prompt, are skipped.

An image generation AI first turns the words of the prompt into a string of numbers. The part that reads the words is CLIP, and it is built from many layers stacked together. Clip skip decides how many layers, counting from the back, are skipped. 1 skips nothing (uses it to the end), while 2 skips the last layer and uses the result of the layer before it. The more layers you skip, the more roughly the meaning of the words is conveyed.

Setting	Meaning	How it is written in ComfyUI
Clip skip 1	Use it up to the last layer	CLIP Set Last Layer is -1
Clip skip 2	Skip the last layer	CLIP Set Last Layer is -2
Clip skip 3 or more	Skip even more (not usually used)	-3 or lower

Clip skip: 2 (an example of a common setting for anime models) / Clip skip: 1 (an example of a common setting for SDXL models)

Should you choose 1 or 2?

Match the value written on the model page. If none is written, compare 1 and 2 with the same seed.

The basic rule for Clip skip is to match the setting used when the model was trained. For anime-style Stable Diffusion 1.5 models, 2 is often recommended, and many SDXL models are said to be used at 1. But this differs by model, so you cannot say "it is SDXL, so it must be 1." If you cannot decide, read the model description (model card), and if nothing is written, compare using the steps below.

Steps for deciding by comparing Clip skip

Look on the model page to see whether a recommended Clip skip is written
If it is, use that value and make one picture
If it is not, keep the seed fixed and make two pictures with Clip skip at 1 and 2
Choose the one that draws the picture as the prompt says (hair style, clothes and background are matched)
Write down the value you chose along with the name of the model

For how to fix the seed and match conditions, see What Is a Seed? Fixed, Random, and Reproducing the Same Image →

To line up 1 and 2 and compare them at once, see How to Use the XYZ Plot (Compare Settings Side by Side) →

Anime girl in a blazer uniform with a red ribbon, an example for checking how well the prompt is followed — 言葉どおりか確かめる絵（1）

Anime girl holding a book in front of library shelves, a reference for comparing how words take effect — 言葉どおりか確かめる絵（1）

Illustration usage rules

Is the place to set it different between tools?

Yes. The item names and the positions of sliders differ between tools.

In AUTOMATIC1111-style screens, you choose the number in the "Clip skip" item inside the settings (or in a box you add to the top of the screen). In ComfyUI you use a node called "CLIP Set Last Layer," where -1 is Clip skip 1 and -2 is Clip skip 2. Initial values and display positions differ between tools, so check the explanation for the screen you use.

Tool	Where to set it	How it is written
AUTOMATIC1111-style	The settings screen (or a box at the top of the screen)	Clip skip: 2
Forge and similar	The settings screen	The item name may differ by version
ComfyUI	The CLIP Set Last Layer node	stop_at_clip_layer is -2

How do you review things when a picture looks wrong? (Common failures and fixes)

First match Clip skip to the model's recommendation, and if that does not fix it, look for another cause.

Common failure	Cause	Fix
The words of the prompt are hard to reflect	The model and Clip skip do not match	Match the value in the model description
The picture is blurry overall and the mood is off	You are using a value different from training	Compare 1 and 2 with the same seed and choose
The face or clothes broke after switching to 2	That model was trained at 1	Go back to 1
It does not become the same picture as someone else's sample	Settings other than Clip skip (sampler, VAE) also differ	Look at the generation info and match all the settings

Change Clip skip from 1 to 2 and compare using the same prompt with the seed kept fixed

To go through every cause when words do not show up in the picture, see Why your prompt is not working (tags ignored) and how to fix it →

For how to choose a model and read its description page, see How to Choose a Stable Diffusion Model (Checkpoint) →

FAQ

Frequently Asked Questions

Does Clip skip work on everything?

No. It is a setting that changes slightly how the prompt is read, so it does not greatly change the composition or art style. Think of it as an adjustment that changes how well words take effect a little.

Is Clip skip 2 required for anime models?

It is not required. It is said that many models recommend 2, but it differs by model. Following the model description page is the surest.

Should I change Clip skip for SDXL models?

In many cases you use it at 1. But some models recommend another value, so check the description page.

How do I set Clip skip to 2 in ComfyUI?

Add a CLIP Set Last Layer node and set stop_at_clip_layer to -2. -1 corresponds to Clip skip 1. The node name may change between versions.