Denoising strength is a slider from 0 to 1 that controls how much an AI model is allowed to change a source image when you use img2img, inpainting, or hires fix. A low value keeps the original almost untouched, a high value reinvents it from scratch, and everything in between trades faithfulness for freedom.
If you have ever loaded a picture into an AI generator, hit generate, and watched it either come out identical to what you started with or turn into something completely different, denoising strength is the dial you were fighting. Anyone editing existing images, upscaling a favorite render, or fixing one small area of an otherwise good NSFW result runs into this setting fast, and getting it wrong is the single most common reason edits look broken.
This guide is for adults 18+ only.
What denoising strength actually is
Denoising strength (sometimes just called “denoise” or “denoising”) is a number between 0 and 1 that tells the model how far it is allowed to travel away from your starting image. It only appears when there is a source image to work from, which means img2img, inpainting, and the hires fix (high-resolution) pass. It does not exist in a plain text-to-image generation, because there is no original to preserve.
Think of it as a leash. At 0, the leash is so short the model cannot move at all and you get your input back unchanged. At 1, there is effectively no leash and the model treats your image as nothing more than a loose suggestion, rebuilding almost every pixel. The whole art of image editing with AI is choosing the right leash length for the job.
To understand why the number matters, it helps to remember how these models build a picture. As explained in how AI image generation works, the model starts from random noise and removes it step by step until an image appears. In img2img, instead of starting from pure noise, the model starts from your source image with a controlled amount of noise mixed in. Denoising strength is literally how much noise gets added before the model starts cleaning it back up. More noise added means more of the original is destroyed and reimagined, so more of the result comes from the prompt rather than the source.
How it works and what it does to your image
The mechanism is simpler than it sounds. A denoise value of 0.3 means the model roughly keeps 70 percent of the original information and rebuilds the other 30 percent according to your prompt and settings. A value of 0.7 flips that, keeping only a small skeleton of the original and generating most of the picture fresh. This is why the visible effect scales so smoothly: the higher you push the number, the less your source image survives.
The effect you actually see falls into a few bands. Very low denoise (roughly 0.1 to 0.25) is for tiny touch-ups and detail passes where the composition, pose, and colors must stay locked. Low to medium (around 0.3 to 0.5) lets the model restyle, clean up, or nudge an image while still respecting its layout. High denoise (0.6 to 0.8 and up) is for major transformations where you only want the model to borrow the rough shapes and go its own way. Push all the way toward 1.0 and you are basically doing text-to-image with a faint hint of the original bleeding through.
For NSFW work specifically, this dial decides whether an edit reads as a subtle refinement of a character you already like or a completely new person. If you found a great pose and body but the face drifted, a low denoise inpaint on just the face fixes it without disturbing the rest. If you want to keep the composition but change the entire outfit and lighting, a higher denoise on the whole frame does it. The number is the difference between “same scene, better” and “new scene entirely.”
Denoising strength by use case
Denoise does not have one correct value. It has a correct value per task, because img2img, inpainting, and hires fix each ask a different question of the model.
In img2img, denoise controls how strongly the model restyles your input. If you are turning a rough sketch or a real photo reference into a rendered image, you often need a higher denoise (0.6 or above) so the model has enough freedom to actually apply the new style. If you just want to clean up an existing AI render or shift its mood, a lower value keeps the composition you already liked.
In inpainting, denoise controls how much the masked patch changes relative to the rest. Fixing a small artifact, like a stray mark or a slightly off hand, wants a lower denoise so the repair blends in. Replacing an entire region with new content wants a higher denoise so the model actually generates something new inside the mask instead of lightly editing what was there.
In hires fix and upscaling, denoise controls how much new detail the model invents while enlarging. A common range here is deliberately low, often around 0.2 to 0.4, because you want sharper skin texture, hair strands, and fabric detail without the model drifting away from the composition you already approved at the smaller size. Push denoise too high during a hires pass and faces mutate, extra fingers appear, and the whole picture “drifts” into something you did not ask for. The exact sweet spot depends on the model and the upscaler, so treat these as starting points and adjust.
Because the right value is so task-dependent, always ask yourself one question before setting it: how much of this image do I want to keep? The answer maps directly to the slider.
Low versus high denoise at a glance
| Denoise range | How much changes | Keeps composition | Best for |
|---|---|---|---|
| 0.1 to 0.25 (very low) | Barely any, surface only | Almost perfectly | Detail passes, tiny artifact cleanup, gentle hires |
| 0.3 to 0.45 (low) | Light restyle, minor edits | Yes, mostly intact | Color and mood shifts, subtle inpaint repairs |
| 0.5 to 0.6 (medium) | Noticeable reinterpretation | Roughly, layout shifts a bit | Style changes, moderate edits, sketch to render |
| 0.65 to 0.8 (high) | Heavy, most pixels new | Loosely, expect drift | Big transformations, replacing whole regions |
| 0.85 to 1.0 (very high) | Near total, original as hint | No, treat as fresh gen | Using the source only as loose inspiration |
Read this as a map, not a rulebook. Different checkpoints and different samplers respond slightly differently, and a distilled or turbo model may need lower values than a standard one to reach the same amount of change. When in doubt, generate the same edit at three values (say 0.35, 0.5, and 0.65) and compare.
How to set it in your tool
In Automatic1111 and Forge, the denoising strength slider sits right below the image drop zone on the img2img and inpaint tabs, and again inside the hires fix panel on the text-to-image tab. It is one slider, clearly labeled, and defaults tend to sit around 0.7 for img2img and much lower for hires fix.
In ComfyUI, denoise is a numeric input on the KSampler node. When you build a text-to-image graph it is set to 1.0 by default (full generation from noise), and when you build an img2img or inpaint graph you lower it to control how much the latent image changes. Because ComfyUI exposes it as a raw node value, it is easy to forget you left it at 1.0 and wonder why your img2img output ignores the source entirely.
Hosted and beginner tools sometimes hide this behind a friendlier label like “edit strength,” “variation,” or “creativity,” but it is the same idea underneath. If a slider promises “more like the original” on one end and “more creative” on the other, that is denoising strength wearing a costume. The beginner settings overview walks through where these live in simpler interfaces.
What to watch for
The classic mistake is running a hires fix or upscale at too high a denoise. You approved a great composition at 768 pixels, sent it to hires at 0.7, and got back a picture with a different face, warped hands, and an extra limb. The fix is almost always to drop denoise into the low range so the enlargement adds detail instead of rewriting content. If hands specifically keep breaking, pair a lower denoise with the techniques in the fix hands guide.
The opposite mistake shows up in inpainting: setting denoise too low when you actually want new content. If you mask an area to replace it and set 0.2, the model barely touches the masked region and you get a faint smudge instead of the change you wanted, often with a visible seam where the mask edge sits. Replacement edits usually need medium to high denoise so the model commits to generating something new inside the mask.
A third trap is expecting denoise to fix a composition problem. If the source image is fundamentally off, cranking denoise just rolls the dice on a new image and throws away the one thing img2img was supposed to give you: control over the layout. When you want a genuinely different scene, it is often cleaner to go back to text-to-image, dial in the prompt, and lock a good seed than to abuse a high denoise img2img pass.
Finally, remember that denoise interacts with everything else. Change the denoise and the same prompt, model, and seed will produce a different result, because you changed how much of the image the prompt controls. Move one dial at a time when you are learning what it does.
A mental model that makes it click
If the numbers still feel abstract, try this picture. Imagine your source image printed on paper, and denoise as how much of that paper you erase before redrawing. At 0.2 you erase a thin top layer and redraw only the surface, so the picture underneath still shows through almost completely. At 0.5 you erase about half, redrawing enough to change the style while the underlying shapes still guide your hand. At 0.8 you erase most of it and the faint outline that remains is barely a suggestion. This is why low denoise is faithful and high denoise is free: the number is quite literally how much of the original you throw away before the model rebuilds.
This model also explains why denoise and steps are different things that people confuse. Steps are how many passes the model takes while cleaning up the noise, which affects finish and coherence. Denoise is how much noise there was to clean up in the first place, which affects how far the result can travel from the source. You can run high steps at low denoise (a careful, faithful refine) or low steps at high denoise (a fast, loose reinvention). They are separate dials that happen to work on the same process.
Denoise across a real workflow
In practice, denoise is not a value you set once, it is a value you move as an image progresses through stages. A common NSFW workflow looks like this. First you generate a base in text-to-image, where denoise does not apply because there is no source. Then you might send a promising result to img2img at a medium denoise to restyle or refine the overall look. Next you inpaint problem areas, using a lower denoise for small repairs and a higher denoise where you want to replace a region entirely. Finally you run a hires or upscale pass at a low denoise to add detail without disturbing the composition you have carefully built.
Notice how the value trends downward as the image gets closer to done. Early on you want freedom to change things, so denoise is higher. Late in the process you have an image you like and mostly want to protect, so denoise drops. Reading your own denoise value as a signal of “how committed am I to this composition” is a useful habit. If you catch yourself running a late-stage polish at high denoise, that is usually a mistake waiting to happen.
When denoise matters most for NSFW
Denoise earns its keep in three NSFW situations especially. The first is rescuing a nearly-perfect render: the pose and body are great but one detail is wrong, and a low-denoise inpaint fixes the detail without gambling the whole image on a re-roll. The second is style conversion, where you take a base composition and push it toward a different rendering at medium denoise while keeping the framing you liked. The third is detail enhancement, where a low-denoise upscale adds convincing skin, hair, and fabric texture that a smaller generation could not resolve.
In all three, the through-line is control. Text-to-image gives you a fresh roll of the dice each time, while denoise-based editing lets you keep what works and change only what does not. That is the entire reason these dials exist, and why learning to read the denoise slider is one of the highest-leverage skills for anyone doing serious editing rather than pure prompting. When you combine a locked composition, a modest denoise, and a good sampler, you move from generating images to actually directing them.
How to find your own sweet spot
Because the right denoise depends on your model, your upscaler, and the exact edit, the fastest way to learn is a short sweep rather than guessing. Pick the image you want to edit, hold everything else fixed, and generate the same edit at three or four denoise values across the range you suspect, for example 0.3, 0.45, 0.6, and 0.75. Lay them side by side and you will immediately see where “too little changed” turns into “too much changed” for that particular task. That boundary is your sweet spot, and it tends to be stable for similar edits on the same model, so you only have to find it once per workflow.
Keep a rough memory of the values that work for you. Many people settle into personal defaults like a low value for every hires pass, a medium value for restyles, and a higher value for region replacement, then only sweep when something looks off. This turns denoise from a mysterious slider into a small set of trusted numbers, which is exactly how experienced users treat it. The setting rewards a little deliberate experimentation early, and then quietly does its job for a long time after.
The short version
Denoising strength is the “how much do I want to change this” dial for any AI edit that starts from an existing image. Low values (around 0.2 to 0.4) preserve your original for detail work, upscaling, and small fixes. High values (0.6 and up) hand the model freedom to transform, replace, and reinvent. The single sanest default to remember: use low denoise for hires fix and upscaling so your approved composition survives, use medium denoise for restyling in img2img, and use medium to high denoise for inpainting when you actually want new content inside the mask. When an edit looks wrong, denoise is the first slider to check, and moving it half a step in either direction usually fixes the problem faster than rewriting the prompt.
Frequently asked questions
What is a good denoising strength for img2img?
It depends on how much you want to change. For a light restyle that keeps the original composition, try 0.3 to 0.45. For a moderate style change, 0.5 to 0.6 works well. For turning a rough sketch or reference into a fully rendered image, you usually need 0.6 or higher so the model has enough freedom. There is no single universal number, so generate at two or three values and compare.
What denoising strength should I use for hires fix or upscaling?
Keep it low, commonly around 0.2 to 0.4, depending on the model and upscaler. A low value lets the model add sharper skin, hair, and fabric detail while enlarging without drifting away from the composition you already approved. Push it too high during a hires pass and faces mutate, hands break, and extra limbs appear.
Why does my img2img output look identical to the source?
Your denoising strength is set too low, so the model is only allowed to change a tiny fraction of the image. Raise the value into the medium range (0.5 or so) to give it enough room to apply your prompt. In ComfyUI, also check that the KSampler denoise value is not accidentally left extremely low.
Why does my img2img output ignore the source completely?
Your denoising strength is too high, likely near 1.0, which tells the model to treat the source as a faint hint and generate almost everything fresh. Lower it into the 0.3 to 0.6 range so the original composition survives. In ComfyUI the KSampler defaults to 1.0, which is a common cause of this.
Does denoising strength exist in text-to-image?
Not as a separate control, because there is no source image to preserve. A pure text-to-image generation effectively runs at full denoise since it starts from random noise. Denoising strength only appears when you feed in an existing image through img2img, inpainting, or a hires fix pass.
What denoising strength is best for inpainting?
It depends on the goal. To repair a small artifact and blend it in, use a lower value so the patch matches its surroundings. To replace a whole region with new content, use a medium to high value so the model actually generates something new inside the mask instead of lightly editing what was already there.
Why do I get seams around my inpainted area?
Often the denoise is too low for the change you wanted, so the patch does not fully commit and a visible edge remains where the mask sits. Raising denoise usually helps, as can adjusting the mask blur or padding in your tool. If you are only doing a tiny fix, a low denoise with a softer mask edge blends better.
Does changing denoising strength change the image even with the same seed?
Yes. Denoise controls how much of the picture the prompt generates versus how much the source contributes, so changing it changes the result even if the seed, prompt, and model stay the same. When you are learning, adjust denoise on its own and keep everything else fixed so you can see exactly what it does.
