Three dials sit under your prompt box. Most people never touch them, and then blame the prompt.
Three dials sit under your prompt box. Most people never touch them, and then blame the prompt.
The three settings everyone leaves on default, and what they are actually doing to your image
Here is a thing I see constantly. Someone spends forty minutes perfecting a prompt, rewrites it nine times, adds weights, tunes the negatives, and gets a muddy, oversaturated, slightly melted picture. Then they blame the prompt. Or the model. Or the seed.
It was CFG 18. It was always CFG 18.
We have written a lot about what goes in the prompt box on this site. We have barely talked about the three controls sitting right underneath it, which is a strange gap, because those controls decide whether your beautiful prompt ever had a chance. Let us fix that. No math, no papers, just what each dial does and where to put it.
You need one mental model and then everything else clicks. When you generate an image, the model does not paint. It starts with a square of pure random noise, like TV static, and then removes a bit of that noise, over and over, guessing what should be underneath. After enough passes, a picture is left behind. That is it. That is diffusion. It is sculpture, not painting: the image was always in the marble, and the model is chipping noise away until it shows.
Now the three dials make sense immediately:
That is the whole settings panel. Method, passes, pressure.
Steps is the easiest one, so let us clear it first and move on.
More steps means more denoising passes, which sounds like it should mean more detail forever. It does not. What actually happens is that the image converges. The big shapes resolve in the first handful of passes, the details fill in over the next batch, and then it plateaus, and further steps just nudge pixels around while you pay full price in generation time.
For most modern models and most samplers, useful convergence lands somewhere around 20 to 30 steps. Below roughly 15 you tend to get soft, undercooked, vaguely smeared results because the model genuinely ran out of passes before it finished. Above 30 or so you are usually buying nothing.
The honest test takes two minutes and beats any number I can give you: lock your seed, lock your prompt, and generate the same image at 15, 20, 25, 30, 40, and 50 steps. Look at where it stops changing. That number is your setting for that model. It is not the same for every model, which is exactly why nobody can hand you a universal answer.
The exception that matters: ancestral samplers, the ones with an "a" in the name like Euler a, inject a little fresh noise at every step. They never fully converge, so more steps does not settle the image, it keeps changing it. That is not a bug, and we will come back to it.
CFG stands for classifier-free guidance, and you can forget that immediately. Here is what it means in practice: it is a tug of war between what you asked for and what the model thinks looks right.
Low CFG hands the model the wheel. It will make something coherent and pretty and it will quietly ignore half your prompt. High CFG hands you the wheel. The model obeys your words harder and harder, and past a point it will obey them right off a cliff, because following your prompt to the letter and making a believable image are two different goals, and you just told it to stop caring about the second one.
What too-high CFG looks like, so you can recognize it on sight: colors go radioactive, contrast slams to the extremes, edges get crunchy and haloed, skin turns to plastic or orange, and everything takes on that deep-fried, oversharpened quality. That is not a style. That is the image tearing.
| CFG range | What you get | When to use it |
|---|---|---|
| 1 to 3 | Model does mostly what it wants, prompt is a loose suggestion, results are soft and dreamy | Abstract work, or distilled models that require it |
| 4 to 6 | Gentle guidance, natural look, some prompt elements drift | Photoreal work where you want the model's instincts |
| 7 to 9 | The reliable middle. Prompt is followed, image still looks like an image | Start here for almost everything |
| 10 to 13 | Aggressive adherence, contrast climbing, starting to look processed | When the model keeps ignoring a specific element |
| 14 and up | Burnt, oversaturated, crunchy, frequently deformed | Essentially never on a standard model |
If the model is ignoring part of your prompt, your instinct will be to crank CFG. Resist it. Cranking CFG to force one word to land will wreck the other forty. Fix it in the prompt first, with word order and weighting, which we broke down in the prompt anatomy guide. Use CFG as a nudge, not a hammer.
Big asterisk for distilled models: Flux and the various Turbo, Lightning, and LCM variants are distilled, meaning guidance is partly baked in during training. They want very low guidance values, often in the 1 to 3.5 range, and they run in a handful of steps rather than 20 to 30. Feed one of those a CFG of 7 and you will get garbage and conclude the model is broken. It is not. You just brought the wrong ruler.
Your dropdown probably has thirty entries. You need roughly three, and the meaningful split is simpler than the names suggest.
These settle. Give them enough steps and the image locks in and stops changing. The workhorse family here is DPM++ 2M, usually paired with a Karras schedule, and it is the default recommendation for a reason: it is fast, it is stable, and it gets a clean result in around 20 to 30 steps. If you want a reproducible image you can iterate on, you want a converging sampler.
These are the ones with an "a" in the name, like Euler a and DPM++ 2S a, and they add a splash of new noise every step. Because of that they never truly settle, and the same seed at 25 steps and 26 steps can give you meaningfully different images. That sounds like a defect and it is actually a feature: ancestral samplers are more creative, more varied, and better at rescuing a boring composition. They are just terrible when you need consistency, because they fight you.
Plain Euler, no "a", is the honest simple one. Fast, predictable, a little less refined than DPM++ 2M, and completely fine. It is a good baseline when you are testing something else and want the sampler to stop being a variable.
| Sampler | Type | Character | Reach for it when |
|---|---|---|---|
| DPM++ 2M Karras | Converging | Clean, fast, reliable | Default. Most work, most of the time |
| Euler a | Ancestral | Creative, varied, unsettled | Exploring ideas, breaking out of a rut |
| Euler | Converging | Simple, predictable | Baseline testing, removing variables |
| DPM++ SDE Karras | Stochastic | Detailed, slower | Final renders when time does not matter |
Here is the part that saves you a week: sampler choice matters far less than you think. The gap between DPM++ 2M Karras and Euler a on a good prompt is real but small. The gap between CFG 7 and CFG 18 is catastrophic. People spend hours sampler-shopping to fix problems that live entirely in the CFG box.
Start here, on a standard non-distilled model, and adjust from evidence rather than vibes:
That combination is boring and it is right about ninety percent of the time. From there, change exactly one dial at a time with a locked seed. If you change the sampler and the CFG and the steps together and the image improves, you have learned nothing, because you cannot say which change did it. This is the same discipline behind locking seeds for controlled variation, and it is the difference between building a workflow and rolling dice.
The honest counterpoint, since I have been prescriptive: these numbers are conventions, not physics. A heavily fine-tuned checkpoint may have been trained toward a different sweet spot, and its model card usually says so. Read it. The card beats the blog post, including this one.
Your image starts as static and the model chips noise away until a picture is left. The sampler is the chipping method, steps is how many swings it gets, and CFG is how hard you are shouting instructions while it works. Shout too loud and it panics, which is what burnt, orange, crunchy output actually is. Give it too few swings and it never finishes. Pick DPM++ 2M Karras, 25 steps, CFG 7, then move one dial at a time on a locked seed. And if you are on Flux or a Turbo variant, throw the CFG advice out entirely and go find the low number the model was built for.
Next time an image comes out looking deep-fried, do not rewrite the prompt. Look under it.