What the process looks like from the inside
Generation runs backwards. The model starts from noise and denoises it in stages, and at each stage it predicts what the picture underneath would look like with a little less noise. Text to image starts from pure noise, so there is nothing to preserve. Image to image starts from your picture with noise added on top, and how much noise gets added is exactly what denoising strength means.
At 1.0 your input is buried completely and only the prompt survives. At 0.3 a faint version of the original is still visible through the noise, so the model rebuilds around the shapes it can still see. Every practical decision about image to image is a decision about how much of the source you want to remain legible to the model.
Picking a strength
| Strength | What survives | Use for |
|---|---|---|
| 0.15 to 0.3 | Everything, subtly altered | Grading, texture, film grain, small cleanups |
| 0.35 to 0.55 | Composition and pose | Restyling, changing medium, repainting |
| 0.6 to 0.75 | Rough shapes only | Sketch to render, loose layout reference |
| 0.8 to 1.0 | Almost nothing | New image with a hint of the old |
The failure modes sit at both ends and look nothing alike. Too low and the output is your input again, which people misread as the prompt being ignored. Too high and identity goes: the face is a different person, the product no longer matches the packaging, and the pose you were trying to keep has drifted.
Steps, and the trap of multiplying them
Steps are how finely the path from noise back to image is sliced. More slices means each one is a smaller, safer jump, up to the point where the result has converged and further slices change nothing you can see.
The trap is the interaction. In most image to image implementations the model only runs the fraction of the schedule that strength asks for, so a low strength job silently gets very few real denoise passes. Twenty steps at 0.25 strength is five effective passes, which is why gentle edits so often look soft and unresolved. The fix is not more strength, which would change more than you wanted, but more steps: 40 to 60 at low strength costs little and resolves properly.
What no amount of either will fix
Neither dial adds knowledge or corrects layout. If the model cannot draw a convincing hand it will denoise its way to a bad hand at every strength and every step count. If the composition is wrong, the fix is a different seed, a structural reference, or an inpaint pass on the offending region.
One habit that saves a lot of credits: when a result is close but not right, decide first whether you want more change or cleaner change. More change is strength. Cleaner change is steps. Reaching for the wrong one is the most common way an editing session stalls.