Clip skip is a setting that controls how many of the final layers of the CLIP text encoder the model skips when it turns your prompt into guidance. Clip skip 1 uses the full text understanding, while clip skip 2 stops one layer early. Many anime, Illustrious, and Pony-lineage NSFW models were trained expecting clip skip 2, so it gives them cleaner, more on-model results.
If you have ever loaded a popular anime NSFW checkpoint, followed the prompt examples exactly, and still gotten results that looked slightly off compared to everyone else, clip skip is very often the missing setting. It is a small, one-click change that beginners rarely know about, and on the right model it makes a real difference.
This guide is for adults 18+ only.
What clip skip actually is
Before the model can draw anything, it has to understand your prompt. That job belongs to the text encoder, called CLIP, which reads your words and converts them into the numerical guidance the image model follows, part of the pipeline covered in how AI image generation works. CLIP processes your text through a stack of layers, and each layer refines the representation a little further, from broad meaning toward fine detail.
Clip skip decides how deep into that stack the model goes before it hands the result to the image generator. Clip skip 1 means “use the output of the final layer,” the full, complete text understanding. Clip skip 2 means “stop one layer short and use that instead.” Higher numbers skip more layers, but in practice almost everyone only ever uses 1 or 2, because those are the two values models are actually trained around.
Why would stopping early ever help? Because a model learns to interpret prompts based on whichever layer it was trained to read from. If a checkpoint was trained using the second-to-last layer, then generating with the full final layer feeds it a slightly different signal than it expects, and results drift off-model. Matching clip skip to what the model was trained on is the whole point.
What it does and the effect on your image
The effect of clip skip is on how the model reads your prompt, which in turn changes the look and accuracy of the output. On a model trained for clip skip 2, using clip skip 2 gives cleaner, more coherent, more on-style images that match the model’s example gallery. Using clip skip 1 on that same model can produce results that feel slightly muddy, less accurate to the tags, or just not quite as crisp as they should be.
It is important to be clear that this is not a universal “quality” slider you crank for better images. It is a matching setting. On a realistic photo model trained for clip skip 1, forcing clip skip 2 usually makes things worse, not better, because now you are the one feeding it an unexpected signal. The right value is entirely about what the model expects.
For NSFW work this matters most in the anime and illustrated space, where the Illustrious and Pony lineages dominate. These families are overwhelmingly trained expecting clip skip 2, which is why their community example prompts almost always assume it. If your anime NSFW output looks off despite a good prompt, clip skip is one of the first things to check. The best Illustrious checkpoints roundup and the anime NSFW generators guide both live in this world where clip skip 2 is the norm.
Anime and Pony models versus realistic models
The split is fairly clean and easy to remember. Anime, illustrated, and Pony or Illustrious-derived NSFW models generally want clip skip 2, because that lineage was trained that way and the tag-based prompting these models use behaves best at that setting. Realistic SD 1.5 and SDXL photo models generally want clip skip 1, because they were typically trained on the full final layer.
This is why changing model families and forgetting to change clip skip is such a common mistake. You switch from an anime checkpoint at clip skip 2 to a realistic one, leave the setting where it was, and the realistic model now underperforms because it is getting a signal it was not trained for. The habit to build is to check clip skip whenever you change models, the same way you would check the recommended VAE or sampler.
When a model page states a recommended clip skip, follow it, because the author knows how they trained it. When it does not say, the family rule of thumb (anime or Pony to 2, realistic to 1) is a safe default.
Clip skip 1 versus 2 compared
| Setting | Uses which CLIP output | Best for | Typical effect if mismatched |
|---|---|---|---|
| Clip skip 1 | Full final layer | Realistic SD 1.5 and SDXL photo models | Anime/Pony models look slightly muddy or off-tag |
| Clip skip 2 | One layer early | Anime, Illustrious, Pony-lineage models | Realistic models can lose accuracy and crispness |
The grid makes the rule concrete: there is no globally better value, only a better value for the model you are running. Match it to the model family, or to the model page’s explicit recommendation, and move on.
How to set clip skip
In Automatic1111 and Forge, clip skip is usually a setting you enable in the options and then add to the top toolbar as a slider, so you can switch it per generation. It is a simple integer, and you will almost always leave it on 1 or 2. Once it is on the toolbar, changing it takes one click whenever you swap model families.
In ComfyUI, clip skip is handled with a “CLIP Set Last Layer” node, where the value is expressed as a negative number (for example -1 corresponds to the full output and -2 to skipping one layer). It is wired into your text encoding, and once you understand the mapping it behaves exactly like the slider in other tools.
Many hosted and app-based generators either fix clip skip to the correct value for their models automatically or hide it entirely, so this is mostly a local-generation concern. The beginner settings overview points out where less common settings like this live, and the CFG and sampler tuning guide covers the neighboring dials you will usually adjust in the same session.
What to watch for
The number one trap is forgetting to switch clip skip when you change model families. Because it is a “set and forget” toolbar setting, it quietly stays where you last left it, and then your next model runs at the wrong value without any obvious error. If a new model looks worse than its example images, check clip skip before you start rewriting the prompt.
The second trap is blaming the prompt for a clip skip mismatch. When output looks slightly muddy, less accurate to your tags, or just not as clean as the model’s gallery, people often assume their prompt is weak and start adding quality tags. On an anime model that actually wants clip skip 2, the real fix is the one setting, not a longer prompt. Try flipping clip skip first, it is a five-second test.
The third thing to keep in mind is that clip skip is not a universal quality booster. Setting a realistic model to clip skip 2 in hopes of “better” images usually degrades them. Resist the urge to treat it as a magic dial. It is a compatibility setting, and higher is not better, matched is better.
Finally, do not confuse clip skip effects with other issues. Dull colors point to a VAE, blur can point to sampling or upscaling, and prompt adherence points to guidance. Clip skip specifically changes how your text is interpreted, so its symptoms are about on-model accuracy and tag fidelity rather than color or sharpness.
Why a whole model lineage settled on clip skip 2
It can seem arbitrary that so many anime and Pony-derived NSFW models expect clip skip 2, so it helps to know the history in plain terms. Early in the illustrated-model scene, a very influential set of anime models was trained using the second-to-last CLIP layer rather than the final one, and those models became the foundation that later checkpoints were built on and merged from. Because so much of the lineage traces back to that choice, the expectation propagated forward, and today the practical reality is simply that this family of models reads prompts best at clip skip 2.
The Illustrious and Pony lineages that dominate current anime NSFW generation inherit this convention, which is why their example prompts, community settings, and recommended workflows almost universally assume clip skip 2. You do not need to understand the training theory to benefit from it. You just need to know that if you are in the anime or Pony world, clip skip 2 is the home setting, and departing from it is usually a mistake rather than a creative choice.
Realistic photo models come from a different training tradition and generally use the full final layer, which is clip skip 1. This is the root of the whole split: two lineages, two training conventions, two correct settings. Once you frame it as history rather than a quality dial, the rule of thumb stops feeling random and starts feeling obvious.
The practical upshot is that you can usually guess the right clip skip from the kind of model you are loading, before you even read its page. Downloading a cute or stylized anime NSFW model from the Illustrious or Pony side of the world? Assume clip skip 2 until told otherwise. Downloading a photoreal model aimed at lifelike bodies and skin? Assume clip skip 1. Then let the model’s own page confirm or override that guess. This is the same instinct you build for other settings: you learn the family conventions, apply them by default, and only deviate when an author gives a specific reason. The best anime NSFW generators guide lives almost entirely in clip-skip-2 territory, while realistic-focused roundups sit on the clip-skip-1 side, so the split maps cleanly onto the kind of content you are making.
Clip skip as part of a model’s setup profile
The useful way to think about clip skip is as one item in a small profile of settings that travel with a model. When you download a new checkpoint, it typically comes with a few recommendations from its author: a suggested sampler and step count, a guidance range, sometimes a specific VAE, and often a clip skip value. Treating these as a set, and applying them together when you load the model, saves you from the classic situation where output looks off because one leftover setting from the previous model is now wrong.
Building this habit is easy. Whenever you switch models, do a quick pass: correct clip skip, correct VAE, sampler and steps in the recommended range, guidance sensible for the family. It takes seconds and eliminates a whole category of “why does this look worse than the examples” confusion. The beginner settings overview and the CFG and sampler tuning guide cover the neighboring dials, and clip skip slots naturally into the same checklist.
When you can ignore clip skip entirely
For completeness, it is worth saying when this setting is not your concern. If you generate exclusively on realistic SD 1.5 or SDXL photo models, clip skip 1 is almost always correct and you rarely need to touch it. If you use a hosted tool or app that fixes clip skip internally for its models, the decision is made for you and there is nothing to change. And if you only ever run one model, you set clip skip once to match it and forget it. The setting only becomes something you actively manage when you move between model families, especially in and out of the anime and Pony space, which is precisely where forgetting it causes the most visible drop in quality.
How to tell a clip skip problem from a prompt problem
Because a clip skip mismatch and a weak prompt can both produce underwhelming images, it helps to know how to tell them apart quickly. The signature of a clip skip issue is that the model feels like it is not quite understanding your tags: on an anime checkpoint the results look a little muddy, characters and styles that the model clearly knows come out generic, and the output does not match the crispness of the model’s own gallery even though your prompt matches its examples. That “close but consistently off” feeling across every generation points at a setting, not at your wording.
The fastest confirmation is a two-image test. Keep the seed, prompt, and everything else fixed, generate once at clip skip 1 and once at clip skip 2, and compare. If the anime model snaps into cleaner, more on-model output at clip skip 2, you have found the problem in about ten seconds, and you never had to touch the prompt. Contrast this with a genuine prompt weakness, where the fix is adding detail or structure to your wording, an area the prompt formula guide covers. Training yourself to run the clip skip test before rewriting the prompt saves a lot of wasted effort on anime and Pony models, and it reinforces the habit of treating clip skip as part of the model’s setup rather than an afterthought.
The short version
Clip skip controls how many final layers of the CLIP text encoder the model skips when reading your prompt, and it is a matching setting, not a quality slider. The rule of thumb is easy: anime, Illustrious, and Pony-lineage NSFW models want clip skip 2, realistic SD 1.5 and SDXL photo models want clip skip 1, and when a model page states a value, follow it. The sane default: put clip skip on your toolbar, set it to 2 for anime and Pony models and 1 for realistic ones, and get in the habit of checking it every time you change model families. It is one click, and on the right model it is the difference between output that looks off and output that matches the gallery.
Frequently asked questions
What is clip skip in AI image generation?
Clip skip controls how many of the final layers of the CLIP text encoder the model skips when it turns your prompt into guidance. Clip skip 1 uses the full text understanding, while clip skip 2 stops one layer early. It is a matching setting: the right value depends on how the model you are using was trained, not on a universal quality scale.
Should I use clip skip 1 or 2?
Use clip skip 2 for anime, Illustrious, and Pony-lineage NSFW models, since that family is generally trained expecting it. Use clip skip 1 for realistic SD 1.5 and SDXL photo models, which are usually trained on the full final layer. If the model’s own page recommends a specific value, follow that, because the author knows how they trained it.
Why do anime and Pony models want clip skip 2?
Because that lineage was overwhelmingly trained using the second-to-last CLIP layer, so generating at clip skip 2 feeds the model the signal it expects and produces cleaner, more on-style, more tag-accurate results. Their community example prompts almost always assume clip skip 2, which is why following those examples at clip skip 1 can look slightly off.
Does clip skip improve image quality?
Not on its own. Clip skip is a compatibility setting, not a quality slider. Matching it to what the model expects improves results, but forcing clip skip 2 on a realistic model that wants clip skip 1 usually makes output worse. Higher is not better, matched is better. Set it to the value the model family or model page calls for.
How do I set clip skip in Automatic1111 or Forge?
Enable it in the settings and add it to the top toolbar as a slider so you can change it per generation. It is a simple integer you will almost always leave on 1 or 2. Once it is on the toolbar, switching it takes one click, which makes it easy to adjust whenever you change model families.
How does clip skip work in ComfyUI?
ComfyUI uses a CLIP Set Last Layer node, where the value is expressed as a negative number. Roughly, -1 corresponds to using the full output and -2 to skipping one layer, which matches clip skip 1 and clip skip 2 in other tools. You wire it into your text encoding, and it behaves the same as the slider elsewhere once you learn the mapping.
Why does my anime model look off even with a good prompt?
A very common cause is clip skip being set to 1 when the model expects 2. On an anime or Pony-lineage checkpoint, that mismatch makes output look slightly muddy or less accurate to your tags, even with a strong prompt. Before rewriting the prompt, flip clip skip to 2 and regenerate, since it is a quick test that often fixes it.
Do I need to change clip skip when I switch models?
Often yes. Clip skip is a set-and-forget toolbar setting, so it stays where you last left it and can silently be wrong after you change model families. If you go from an anime model at clip skip 2 to a realistic one that wants clip skip 1, update it. Build the habit of checking clip skip every time you swap model families.
