Denoising strength is a setting in AI image generation that controls how much an existing image changes during processing. It ranges from 0 to 1, where 0 preserves the original image completely and 1 replaces it with something entirely new. You’ll encounter this setting whenever you use image-to-image generation, inpainting, or upscaling in tools like Stable Diffusion.
How Denoising Strength Works
Diffusion models generate images by starting with pure noise and gradually cleaning it up, step by step, until a recognizable image emerges. When you work with an existing image instead of generating from scratch, the model needs to know how far back in that noise process to start. That’s what denoising strength controls.
At a denoising strength of 0, no noise is added to your input image, so nothing changes. At 0.5, the model mixes a moderate amount of noise into the image before beginning its cleanup process, preserving the general composition but allowing significant changes to details. At 1.0, the input image is completely replaced with random noise, and the model generates something from scratch that may bear no resemblance to your original.
Think of it like restoring a painting. A low value is like lightly retouching a few brushstrokes. A high value is like painting over most of the canvas. At maximum, you’ve primed over everything and started a new painting entirely.
Image-to-Image Generation
The most common place you’ll use denoising strength is in image-to-image (img2img) workflows, where you feed in a reference image along with a text prompt. The strength value determines the balance between keeping your original and letting the AI create something new. A low setting keeps the colors, composition, and structure of your source image mostly intact while making subtle refinements. A high setting gives the model freedom to explore more creative transformations, using your image as a loose starting point rather than a strict guide.
This makes denoising strength useful for style transfers, where you might want to keep the layout of a photograph but render it in a painterly style. A value around 0.4 to 0.6 typically preserves recognizable elements of the source while allowing meaningful stylistic changes. Going above 0.7 often produces results where the original image is barely recognizable.
Inpainting Settings
In inpainting, you mask a specific region of an image and ask the model to regenerate just that area. Denoising strength controls how dramatically the masked region changes. At low values, the inpainted area blends gently with its surroundings and makes only minor adjustments. At high values, the model generates entirely new content inside the mask.
A denoising strength of 0.75 is a reliable starting point for most inpainting tasks. If the result changes too much or doesn’t blend well with the rest of the image, dial it down. If the model isn’t making enough of a change, push it higher. Keeping your masked content set to “original” and adjusting denoising strength from there works for the vast majority of inpainting situations.
Upscaling and High-Res Fix
When you upscale an image using latent upscalers (built into tools like Automatic1111’s high-res fix), denoising strength determines how much new detail the model adds versus how faithfully it reproduces the original. This is where the setting gets tricky, because different upscalers respond to the same value very differently.
With standard upscalers like 4x UltraSharp, most users get good results in the 0.2 to 0.4 range. This adds fine details like hair strands, skin texture, and foliage without changing the image’s content. Going much above 0.4 with these upscalers risks altering faces, changing compositions, or introducing unwanted elements.
Latent upscalers are a different story. They tend to produce blurry or artifact-heavy results below 0.5, and many users find they need values of 0.55 to 0.7 to get clean output. The tradeoff is that higher values also change the content more. A denoising strength of 0.5 is roughly the threshold where an upscaled image still closely resembles the original when using latent methods. Testing in the 0.5 to 0.7 range and comparing results is the most reliable approach, since the ideal value shifts depending on the upscaler, the model, and the image itself.
Denoising Strength vs. CFG Scale
These two settings are easy to confuse because they both influence how the final image looks, but they control completely different things. Denoising strength determines how much the original image is preserved or destroyed before generation begins. CFG scale (classifier-free guidance) determines how closely the model follows your text prompt during generation.
A high CFG scale forces the model to conform tightly to whatever words you typed, sometimes at the cost of image quality. A low CFG scale gives the model more creative freedom but may drift from your prompt. CFG scale operates on a range of roughly 1 to 30, with values between 7 and 12 being common defaults.
In practice, these settings interact. If you set a high denoising strength (giving the model lots of room to change the image) but a low CFG scale (giving it freedom to ignore your prompt), the output can veer in unpredictable directions. Conversely, high denoising strength paired with high CFG scale produces images that change dramatically from the original but stick closely to your text description. Understanding that denoising strength controls “how much changes” while CFG scale controls “what guides the change” helps you adjust each one independently.
Practical Starting Points
There’s no single best value for denoising strength because it depends entirely on what you’re trying to do. But these ranges give you a useful starting framework:
- Minor touch-ups and subtle refinements: 0.1 to 0.3. Good for cleaning up small imperfections or adding slight detail without altering the image noticeably.
- Style transfer with preserved composition: 0.4 to 0.6. The image keeps its overall layout and subject matter but gains noticeable stylistic changes.
- Significant transformation: 0.6 to 0.8. The original image serves as a loose guide, but major elements can shift, merge, or disappear.
- Near-complete regeneration: 0.8 to 1.0. Very little of the original survives. At 1.0, the input image has no meaningful influence on the output.
For upscaling specifically, start around 0.3 with non-latent upscalers and around 0.55 with latent upscalers, then adjust based on whether you’re seeing blur, artifacts, or unwanted content changes. For inpainting, 0.75 is a solid default. In all cases, small adjustments of 0.05 to 0.1 can produce noticeably different results, so it’s worth experimenting in small increments rather than making large jumps.