img2img denoising strength, explained
Denoising strength is the single most important slider in Stable Diffusion’s img2img tab, and the most misunderstood. It is a value from 0.0 to 1.0 that decides how far the model is allowed to drift from your source image. Pick the wrong value and you either get an output identical to the input or one that has nothing to do with it. This advisor maps your transformation goal and source image type to a recommended starting range so you stop guessing.
How denoising strength works at a technical level
img2img does not start from pure noise the way text2img does. Instead, the pipeline:
- Takes your source image and encodes it into the model’s latent space.
- Adds Gaussian noise to that latent representation — the amount is controlled by the denoising strength.
- Runs the diffusion sampler for
steps × denoising_strengtheffective steps, denoising from the noised latent back toward a clean image, guided by the prompt and conditioned on the remaining structure from the source.
At 0.0, no noise is added and the output equals the input almost exactly. At 1.0, the latent is fully noised and the pipeline ignores the source almost entirely — equivalent to text2img. Every value between is a dial between “preserve” and “invent.”
The strength scale in practice
| Strength | What survives from source | What changes |
|---|---|---|
| 0.0–0.2 | Nearly everything | Almost nothing; tiny texture shifts |
| 0.25–0.4 | Composition, colours, pose | Fine details, style, texture |
| 0.45–0.6 | Composition, rough shapes | Style, colours, textures, detail |
| 0.65–0.8 | Rough layout, silhouette | Nearly everything else |
| 0.85–1.0 | Almost nothing | Full creative reimagining |
Worked examples by use case
Sketch to polished render (source: rough pencil sketch): The source has form but no detail. Start at 0.65–0.75. The model uses the rough shapes to constrain proportions while inventing all lighting, texture, and colour. Go lower (0.5–0.6) if you want the sketch lines preserved as a structural suggestion.
Style transfer on a photograph (source: photo, target style: oil painting): The subject and composition should stay; the aesthetic should change. Start at 0.5–0.6. Lower if the face identity breaks; higher if it still looks like a photo. Lock the seed and adjust by 0.05 increments.
Inpainting blend (fixing a hand or artifact via masked region): The mask region must blend with the surrounding unmasked pixels. Use 0.3–0.5. Too high and the fix looks copy-pasted in from a different lighting environment; too low and the artifact survives.
Consistent character variation (multiple poses from one base): Start around 0.4–0.55. The model keeps the character’s features while allowing pose and expression changes. If you go higher, the face drifts toward the prompt description rather than the source.
Upscaling with texture enhancement (Hires.fix or img2img upscale): Use 0.35–0.5. This range adds skin pores, fabric weave, and detail without changing the composition. Higher values in upscaling cause the model to creatively reinterpret the image rather than simply enhancing it.
Debugging when results are wrong
- Output looks identical to input: strength is too low, typically below 0.3. Raise it by 0.1 and regenerate.
- Output ignores the source completely: strength is too high. Drop below 0.6 and try again.
- Face or identity breaks: lower strength; consider a strength of 0.4–0.5 and rely more on the prompt for style.
- Fix does not blend with surrounding pixels: reduce strength on the masked region in inpainting; try 0.3–0.45.
Keep the seed fixed when experimenting so that only the slider is changing between generations.