Confused by all the sliders? Here are safe beginner starting values: steps 25, CFG 6, sampler DPM++ 2M Karras, resolution 1024×1024 for SDXL, seed random until you find a keeper, batch 4, hires fix denoise 0.5, and clip skip 2 for anime models. This guide explains what each setting does in plain English and the value to start with.
The settings panel is where most beginners freeze. There are sliders for steps, CFG, sampler, resolution, seed, batch, denoise, and clip skip, and none of them explain themselves. So people either leave everything on default and wonder why results look off, or they crank random sliders and make things worse. Neither is fun.
This page fixes that. For each setting we cover three things: what it actually does, why it matters, and the exact value a beginner should start with. Set these, generate a few images, then adjust one slider at a time so you learn what each one feels like. That is the whole method.
First, the rule that never changes. Subjects must be adult (18+), fictional, and AI-generated. Never real, identifiable people, and never minors. The example prompt at the end carries the safety tokens that belong in every negative prompt. With that locked in, let us go slider by slider.
The beginner settings cheat sheet
| Setting | What it does | Safe beginner value |
|---|---|---|
| Steps | How many refining passes the AI makes | 25 (range 20 to 30) |
| CFG scale | How strictly it follows your prompt | 6 (range 4 to 7) |
| Sampler | The method used to build the image | DPM++ 2M Karras |
| Resolution | Image size in pixels | 1024×1024 (SDXL) |
| Seed | The random starting point | Random, then lock a keeper |
| Batch | How many images per click | 4 |
| Denoise (hires) | How much an image is changed | 0.5 (range 0.4 to 0.6) |
| Clip skip | Affects anime model interpretation | 2 for anime, 1 for photoreal |
Keep this table handy. Now the plain-English explanation of each.
Steps
Steps are how many passes the AI makes refining your image from noise into a finished picture. Think of it like an artist adding layers of detail. Too few steps and the image looks unfinished, grainy, or rough. Too many and you waste time and electricity for almost no visible gain, because the image stops improving after a point.
The sweet spot for most models is 20 to 30 steps. Start at 25. That is enough to get a clean, detailed result without long waits. If your images look noisy or half-baked, nudge steps up a little. If generation feels slow and the extra steps are not helping, bring it down. Beginners often set 50 or 100 steps thinking more is better, but that is mostly wasted time.
A useful way to picture it: the first handful of steps does most of the heavy lifting, roughing in the shapes and composition. The later steps just polish fine detail. Past a certain point, each extra step adds so little that you would struggle to tell two images apart. So 25 is not a compromise, it is genuinely close to the best your image is going to get on that setting alone. Spend your patience on better prompts instead.
CFG scale
CFG scale controls how strictly the AI obeys your prompt. Low CFG lets the model be loose and creative, sometimes ignoring parts of your prompt. High CFG forces it to follow your words closely, but push it too far and the image turns harsh, oversaturated, and fried, with crunchy edges and unnatural skin.
For most models, 4 to 7 is the comfortable range. Start at 6. That follows your prompt well while keeping the image natural. If your results look burnt, contrasty, or overdone, lower the CFG first before changing anything else, since a high CFG is one of the most common causes of ugly beginner output. If the model is ignoring your prompt, nudge CFG up slightly.
Sampler
The sampler is the method the AI uses to build your image step by step. Different samplers produce slightly different looks and run at slightly different speeds, but you do not need to understand the math behind them. As a beginner, just pick one reliable sampler and stick with it while you learn everything else.
DPM++ 2M Karras is a popular, dependable all-rounder that works well across most models. Euler a is another easy, fast choice that gives clean results. Start with DPM++ 2M Karras. Once everything else feels natural, you can experiment with other samplers to see how the look changes, but there is no rush. One good sampler covers you for a long time.
One thing worth knowing so you do not get tripped up: when you change samplers, the same prompt and seed can produce a noticeably different image. That is normal and expected, because the sampler changes the path the model takes to build the picture. So if you are comparing two of your own results, keep the sampler the same and change only one other thing. Otherwise you cannot tell whether the difference came from the sampler or from the setting you actually meant to test.
Resolution
Resolution is the size of your image in pixels. This setting matters more than beginners expect, because each model is trained to compose best at a specific size. Go too small and the image looks soft and lacks detail. Go far too large and you get doubled bodies or stretched anatomy, because the model was never trained to lay out a scene at that size.
For SDXL-based models, which most modern NSFW checkpoints are, 1024×1024 is the native sweet spot. Use 832×1216 for a tall portrait shape or 1216×832 for a wide landscape shape. Start at 1024×1024. Generate at the native size, then enlarge with hires fix or an upscaler for a big, clean final image. Do not type 2048×2048 and expect it to just work.
One practical note on shape. Keep the total pixel count roughly inside the model’s budget even when you change the aspect ratio. The three sizes above all sit in the same comfortable zone. If you stretch one dimension far beyond that, you slide back into doubled-body territory, because you have effectively asked the model to compose a much larger scene. So pick the shape that fits your subject, a tall canvas for a standing pose, a wide one for a reclining scene, but do not balloon the overall size. Match the model to its training, then upscale afterward.
Seed
The seed is the number that sets the random starting point of an image. With a random seed, every generation is completely new, which is exactly what you want while you are exploring ideas. Once you find a composition you love, lock that seed so you can reproduce and refine that exact image instead of rolling the dice again.
Start with a random seed while you hunt for a good result. When one appears, copy its seed and set it as fixed, then tweak the prompt or another setting to improve that specific shot. This single habit turns a lucky image into one you can polish from good to great, and it is how serious creators dial in a final picture.
Most interfaces have a small button, often a recycle or dice icon, that copies the seed of the last image you generated. Learn where it is in your tool early, because it saves you from squinting at long numbers. Many beginners do not realize the seed of every image is recorded, so they lose great results they could have reproduced. Treat the seed like the address of an image: with it you can always find your way back.
Batch
Batch settings control how many images you make per click. There are usually two: batch size, which is how many run at once, and batch count, which is how many rounds. Generating several at a time is smart because AI image-making is a numbers game, and seeing four variations side by side helps you spot the best composition fast.
Start with a batch of 4. That gives you enough variety to compare without overwhelming you or your hardware. If your graphics card is modest, lower batch size and raise batch count instead, so they run in sequence rather than all at once. On a weak card or a phone, generate one or two at a time. The point is to always make a few, not just one.
The difference between the two batch settings trips up a lot of beginners, so here it is plainly. Batch size runs several images simultaneously, which is fast but uses more memory all at once and can crash a smaller graphics card. Batch count runs them one after another, which is slower but gentle on memory. So if you see out-of-memory errors, drop batch size to 1 and raise batch count to 4. You still get four images, just sequentially. Match the split to your hardware and you keep the variety without the crashes.
Denoise (denoising strength)
Denoise controls how much an existing image is changed when the AI works on it again. It mostly matters in two places: img2img and hires fix. Low denoise keeps the result very close to the original. High denoise changes it dramatically, sometimes losing the original composition entirely.
For hires fix, the setting most beginners use, start at 0.5, within a safe range of 0.4 to 0.6. That adds sharpness and detail while keeping your composition intact. Go higher and hires fix may redraw too much, changing faces or poses you liked. Go lower and you get less of the detail boost. For img2img, lower denoise refines a draft gently while higher denoise restyles it more aggressively.
A simple way to remember denoise: think of it as how much freedom you give the AI to repaint. At 0.3 it barely touches the image, perfect for a light polish. At 0.5 it adds real detail while respecting your composition. At 0.7 and above it starts inventing new things, which is great when you want a bold restyle in img2img but risky in hires fix, where you usually want the original kept. When in doubt for hires fix, stay near 0.5 and adjust in small steps. Big jumps in denoise produce big, sometimes unwanted, changes.
Clip skip
Clip skip affects how the model interprets your prompt, and it mainly matters for anime models. Without getting technical, anime checkpoints like Illustrious XL and Pony Diffusion were trained expecting clip skip 2, so setting it there gives cleaner, more accurate anime results. Photoreal models usually want clip skip 1.
The simple rule: clip skip 2 for anime models, clip skip 1 for realistic models. If your anime images look slightly off in style or detail, check that clip skip is set to 2. Many beginners never touch this setting and wonder why their anime output is not as crisp as the examples they have seen. It is a small switch with a real effect on anime quality.
A safe example with all settings filled in
Here is a complete, tasteful setup using every value above. Subjects must be adult (18+), fictional, and AI-generated, never real people and never minors.
Positive: (adult woman, 26 years old:1.2), elegant lingerie, bedroom setting, soft window light, photorealistic, detailed skin, sharp focus
Negative: child, minor, underage, loli, shota, lowres, bad anatomy, extra fingers, blurry, watermark, text
Sampler: DPM++ 2M Karras | Steps: 25 | CFG: 6 | Size: 1024x1024
Seed: random | Batch: 4 | Hires fix denoise: 0.5 | Clip skip: 1 (photoreal)For an anime model, you would switch the prompt to booru-style tags and set clip skip to 2. Everything else stays in roughly the same ranges. That single block is a reliable starting point for almost any NSFW generation, which is why it is worth memorizing the shape of it.
How to actually learn these settings
Reading about sliders only gets you so far. The real learning happens when you change one and watch what it does. Here is the method that works: set every value from the cheat sheet, generate a batch, then change exactly one setting and generate again with the same seed. Compare the two. Now you can see, with your own eyes, what that slider does.
Do that a few times with CFG and with steps, the two you will adjust most, and the whole panel stops being scary. Change one thing at a time, always. Changing five sliders at once teaches you nothing, because you cannot tell which change caused what. Patience and one-variable testing is the entire secret.
If a setting still confuses you by name, our full glossary of NSFW AI terms defines each one with an example. And once your settings are dialed in, how to get better NSFW AI results shows the next round of quality upgrades like ADetailer and stronger prompts. If your images come out soft despite good settings, the blurry image fix guide covers the usual causes.
The best way to make all of this click is to try it. Open our free generator, where sensible defaults are already in place, and start adjusting one slider at a time. You will understand these settings faster in twenty minutes of hands-on play than in an hour of reading. New to the whole process? Start with our beginner generator guide, then come back here whenever a slider puzzles you.
One last reminder. The defaults in this guide are starting points, not laws. Once you understand what each slider does, you will develop your own preferred values for your own style. That is exactly the goal: not to memorize numbers forever, but to understand the dials well enough to tune them with confidence. So set the cheat sheet values, open the generator, and start experimenting today.
And do not feel you must master every setting at once. In practice you will spend most of your time touching just three: the prompt, the seed, and CFG. Steps, sampler, resolution, and clip skip you will set sensibly once and rarely change. So if the panel still feels like a lot, focus your attention on those three and let the rest sit at the cheat sheet defaults. That alone is enough to make consistently good images, and you can deepen your understanding of the other dials gradually, whenever curiosity or a specific problem nudges you to.
Frequently asked questions
What are the safest default settings for a beginner?
Start with steps 25, CFG 6, sampler DPM++ 2M Karras, resolution 1024×1024 for SDXL models, a random seed until you find a keeper, batch 4, hires fix denoise 0.5, and clip skip 2 for anime models or 1 for photoreal. These values produce clean, reliable results across most models and give you a solid base to adjust from one slider at a time.
Does more steps always mean a better image?
No. Steps refine the image, but after about 20 to 30 the gains become tiny while the time and electricity keep rising. Setting steps to 50 or 100 mostly wastes time. Start at 25, which is plenty for a clean result. Only raise it slightly if images look noisy or unfinished, and lower it if generation feels slow without any visible improvement.
Why do my images look fried or oversaturated?
Almost always because your CFG scale is set too high. CFG controls how strictly the AI follows your prompt, and pushing it past about 7 forces the image so hard it turns harsh, contrasty, and unnatural. Lower the CFG to the 4 to 7 range, starting at 6, before changing anything else. That single adjustment fixes most burnt-looking beginner output very quickly.
What sampler should I use as a beginner?
DPM++ 2M Karras is a dependable all-rounder that works well across most models, so start there. Euler a is another easy, fast option that gives clean results. You do not need to understand the math behind samplers. Just pick one reliable choice and stick with it while you learn everything else. You can experiment with others later once the basics feel natural.
What resolution should I generate at?
For SDXL-based models, which most modern NSFW checkpoints are, use the native 1024×1024, or 832×1216 for a tall portrait and 1216×832 for a wide shot. Generate at the native size, then enlarge with hires fix or an upscaler. Avoid typing something huge like 2048×2048 directly, because the model was not trained to compose at that size and you get doubled bodies.
What is hires fix denoise and what value works?
Hires fix denoise controls how much the image is changed while it is enlarged and refined. Start at 0.5, within a safe range of 0.4 to 0.6. That adds sharpness and detail while keeping your composition intact. Go too high and hires fix may redraw faces or poses you liked. Go too low and you get less of the detail boost you wanted from it.
What does clip skip do and when do I change it?
Clip skip affects how the model reads your prompt, and it mainly matters for anime checkpoints, which were trained expecting clip skip 2. Set it to 2 for anime models like Illustrious or Pony, and 1 for realistic photo models. If your anime images look slightly off in style or detail, check that clip skip is set to 2. Photoreal models usually do not need it changed.
Should I generate one image or several at a time?
Several. AI image-making is a numbers game, so making a batch of about 4 lets you compare variations and pick the best composition quickly. If your graphics card is modest, lower the batch size and raise the batch count so they run one after another instead of all at once. On a weak device or phone, generate one or two at a time, but always make more than one.
