liveBook · Manning

Denoising strength

As explained in Local Image Generation

Denoising strength is a parameter that controls how much an input image is altered by diffusion denoising during image-to-image generation or upscaling.

Think of denoising strength as the upscaler’s permission to rewrite your image. It controls how much noise is added before diffusion generates the final result: higher values allow larger changes, while lower values preserve more of the input. section 7.2.2

Use the value as a practical starting point:

  • Set it between 0.0 and 1.0. A value of 0.0 adds no noise; the typical default is 0.7. section 7.2.2
  • Lower it when an upscale changes the original too much. Around 0.5 preserves the source more closely while still adding detail. section 7.2.2 section 7.3.2
  • Do not assume lower is always better. Around 0.3 can smooth details into a blurry or “blobby” result, especially at 4x scale. section 7.2.2
  • When using another checkpoint, reduce the value to moderate its stylistic impact; 0.5 can compromise between dramatic reinterpretation and fidelity. section 7.3.2

Iterating with img2img

A 0.75 setting can push an image toward a more animated, detailed style when the result is sent through img2img again. Set resize to 1 before repeating the process, or the resolution multiplies again; repeated passes can eventually make the image look strange. section 7.3.3

Figure 7.7 Scaling by a factor of 2 with a denoising strength of 0.5.

Compare a 2x upscale at 0.5 with nearby settings to see how preservation and generated detail trade places. section 7.2.2

In a scheduler workflow, denoise also changes the image’s character. A maximum of 1.0 removes the noise completely; small reductions can make the result more illustrative, but larger reductions may rapidly destroy recognizable image quality. section 10.3.3 section 10.5

How Local Image Generation covers it

How deep it goes
Explains the 0.0–1.0 range and visible tradeoffs, but stops short of describing denoising internals.
Examples
Compares 2x and 4x upscaling, checkpoint changes, repeated img2img passes, and scheduler values near 0.9 and 1.0.
What it assumes
Assumes familiarity with image-to-image or upscaling controls, while introducing noise, checkpoints, and scheduler effects through visible results.
What you can do after
Apply it by starting near the default, lowering it for fidelity, and changing it gradually while inspecting detail.
How it connects
Connects denoising strength with image-to-image generation, upscaling, checkpoints, schedulers, and diffusion models.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Local Image Generation. The section links point to the parts of the book it draws on.