Addaly is in open beta. Things will change, and AI answers can be wrong — check anything that matters.

Making Things With AI

Images, video, voice and music — how they work, where they break, who owns them.

Lesson 30 of 848 min

Denoise strength: the number that decides how much survives

One number, and what it means

Every tool that transforms an existing image has a slider called strength, denoise, or image weight. It is the most misunderstood control in the toolkit, and it has an exact mechanical meaning.

Recall that generation walks down a noise schedule from pure static to a clean image. Image-to-image does not start at the top. It takes your input image, encodes it to a latent, adds noise corresponding to some point partway down that schedule, and runs the loop from there.

Strength is where on the schedule you start.

strength 0.2  -> start near the bottom, 20% of the steps run
strength 0.5  -> start halfway
strength 0.9  -> start near the top, almost pure noise
strength 1.0  -> the input is entirely destroyed; this is ordinary generation

Two consequences follow immediately, and both surprise people.

Steps and strength multiply. At 30 steps and strength 0.4, only about 12 steps actually run. This is why an image-to-image pass finishes faster than a fresh generation, and why setting a low strength with few steps produces a barely-changed image no matter what the prompt says.

Steps that actually run, when you asked for thirtyStrength 0.2 — cleanup6Strength 0.35 — restyle,structure held11Strength 0.5 —reinterpretation15Strength 0.7 — the inputis a mood21Strength 1.0 — the inputis discarded30denoising steps that actually runStrength is where on the noise schedule the loop starts, so it multiplies with the step count. At 0.2with thirty requested, six run, and no wording makes a difference the model has no steps left to make.At 1.0 the input is gone and this is ordinary generation.
Steps that actually run, when you asked forthirtyStrength 0.2 — cleanup6Strength 0.35 — restyle, structure held11Strength 0.5 — reinterpretation15Strength 0.7 — the input is a mood21Strength 1.0 — the input is discarded30denoising steps that actually runStrength is where on the noise schedule the loopstarts, so it multiplies with the step count. At 0.2with thirty requested, six run, and no wording makesa difference the model has no steps left to make. At1.0 the input is gone and this is ordinarygeneration.

Strength 1.0 ignores your input completely. People set it high because they want a big change, and at the top of the range they have simply thrown the image away. The band where interesting things happen is narrower than the slider suggests.

What each band is for

Ranges vary by model, and these are close enough to start from:

  • 0.15 to 0.30. Cleanup. Removes JPEG artefacts, unifies lighting, tidies edges. Composition and identity fully preserved. This is the pass that makes a rough composite look like one image.
  • 0.30 to 0.50. Restyling with structure held. A photograph becomes a painting, a sketch becomes a rendering. Faces survive at the lower end and start drifting toward the upper.
  • 0.50 to 0.70. Reinterpretation. The composition is a suggestion; the content is redecided. Useful for turning a rough block-out into a real image.
  • 0.70 and above. The input is a mood, not a plan.

A practical routine, which takes a minute and saves hours: fix the seed, generate the same image-to-image pass at 0.25, 0.4, 0.55 and 0.7, and look at all four. You will develop an intuition for the model's particular scale far faster than by reading anyone's recommended numbers.

The trap that catches everybody

Iterating image-to-image repeatedly degrades the picture in a specific way. Each pass adds noise and denoises, and each denoising pulls the result slightly toward the model's mode. Ten passes at 0.3 do not equal one pass at 3.0; they produce an image that has drifted toward the distribution average, with contrast climbing and detail smoothing on every round.

This is the same mechanism as the distribution pull from the previous module, applied repeatedly. It is also, incidentally, what makes the "keep re-uploading the same image" experiments circulating online converge on a similar glossy face regardless of where they started.

If you need several changes, make them in one pass at a suitable strength, or make them with masks so untouched regions are genuinely untouched.

Where it belongs in a workflow

The professional shape is nearly always the same:

  1. Generate or photograph a base.
  2. Fix the layout by hand — crop, move elements, paint over what is wrong, however crudely. A rough paint-over in GIMP or Krita costs two minutes and does not need to look good.
  3. Run image-to-image at 0.3 to 0.45 to unify it.
  4. Mask and inpaint the specific regions that are still wrong.
  5. Upscale and finish.

Step 2 is the one people skip, and it is the one that makes the rest work. The model is much better at making a bad painting look like a photograph than it is at inventing a composition you have not shown it. A five-year-old's standard of paint-over is genuinely enough, because at strength 0.4 the model is reading shapes and colours, not craftsmanship.

The honest limitation: strength is a single global number. It cannot preserve the face and restyle the background, because it applies to the whole latent. When you find yourself wanting two different strengths in one image, you want masks, which is the next lesson.

The one thing to keep

Image-to-image adds noise to an existing picture and denoises from there, so the strength value sets how far back along the schedule you start and therefore how much of the original can survive.

Before you move on

An image-to-image pass at strength 0.85 returns something almost unrelated to the input picture, though the prompt describes the input accurately. Why?

Pick the one you would defend. Nobody sees your answer.

No ads. No data sale. No public scores on people. Ever.

© 2026 Addaly