Denoise
How much of the starting picture the sampler redraws.
In plain words
denoise is a number from 0 to 1 on the sampler. It says how much noise is laid over the starting latent before sampling, and so how much of it the model redraws.
- 1.0: start from pure noise. Whatever was in the latent is gone. This is text to image.
- 0.5: half-way. The composition and colours of the start stay; texture, faces and detail are redrawn.
- 0.2: only fine detail changes.
With denoise below 1, the sampler runs only the last part of the schedule: at 20 steps and 0.5 it does roughly the last 10.
Where you meet it in a graph
denoiseon KSampler, BasicScheduler and other nodes. KSampler (Advanced) has nodenoise; there the same effect comes fromstart_at_step.- It matters when
latent_imageholds a real picture: from VAE Encode (image to image), from an earlier sampler (hires fix, refining), or with a mask (inpainting).
Values people use
| Task | Typical denoise |
|---|---|
| Text to image | 1.0 |
| Refining with a second model | 0.3 to 0.5 |
| Second pass after upscaling (hires fix) | 0.4 to 0.6 for latent upscale, 0.2 to 0.4 after a model upscale |
| Image to image, keep the picture, change the style | 0.4 to 0.7 |
| Inpainting with an ordinary model | 0.6 to 1.0 |
Typical mistakes
- denoise below 1 on an empty latent: there is nothing to keep, and the picture comes out muddy and unfinished.
- denoise 1 in image to image: the start picture is ignored.
- Too low to change anything, or too high to keep anything: compare two values side by side.
How the app helps
The refining lesson checks denoise of the second sampler between 0.2 and 0.5. The KSampler description warns about denoise below 1 on an empty latent. In Properties, a field with a built-in note has a button What this field does next to its name; it opens the note under the field: what the field does and what to put in it.