Img2img Strength: The Slider That Controls AI Photo Edits

The strength slider decides how much of your original photo survives an AI edit: set it low and you get your photo back with a new mood, set it high and you get a new picture that vaguely remembers yours. Every AI tool that edits one photo into another (image-to-image, or img2img) has this slider, and it matters more than all the others combined. It goes by different names: strength, denoise, creativity, imagination. They all do the same job.

The slider points at one precise moment in the way the model builds a picture. Once you know which moment, the slider becomes predictable, and you can guess the right value before you drag it.

What the slider mechanically does

The model makes pictures by cleaning up speckle. It starts with a screen of pure random static and clears a little of it on each pass. It works down a fixed ladder of steps, and a clean image is what is left at the bottom. (The static is called noise and the cleanup is called denoising. The ladder is the noise schedule, and this kind of model is a diffusion model.) When you generate from words alone (text-to-image), the cleanup always starts at the very top of that ladder, at 100% noise.

Editing a photo this way changes only one thing: where the cleanup begins. The app first turns your photo into the form the model works in. Then it adds noise to it on purpose. That is the same kind of damage the model was trained to undo. Strength decides how much gets added. It pushes your photo partway up the ladder, and the model then works its way back down, guided by your prompt. At strength 0.6, your photo is noised to roughly the 60% point of the ladder. The model does the remaining cleanup itself.

Strength is how deep your photo gets buried before the model digs it back out, rather than a blend or an opacity setting laid over the result afterward. Whatever detail the noise destroys, the model has to invent again from your prompt rather than from your photo. Low strength destroys only fine texture. High strength destroys almost everything except the rough arrangement of light and dark.

What different values look like

The exact result shifts with the model and the picture. Even so, the broad bands hold up well enough to plan around:

The scale does not move in even steps. Going from 0.3 to 0.5 is usually a mild change. Going from 0.6 to 0.8 is often a dramatic one. At those higher settings, the noise eats the medium-sized shapes that made the picture recognizably yours. When in doubt, start around 0.4 and move in increments of 0.1.

There is one wrinkle with the fast models that finish in only a handful of passes (they are usually called turbo models). Strength also sets how many cleanup steps run, because the model only works on the part of the ladder below your starting point. With a four-step model, strength 0.5 leaves roughly two steps of real work. That is usually fine, since these models are built to do a lot in each step. But a very low strength gives the model only a single step to act in. That is why edits below about 0.2 tend to look like nothing happened.

Why faces drift at any strength

The most common disappointment is that everything looks right except the person no longer quite looks like themselves. This happens even at low strength, and the mechanism above explains why. What makes a face yours lives at the scale of millimeters: the exact spacing of the eyes, the precise curve of a jaw. That is the fine detail noise destroys first. Even at 0.25, part of that layer is erased, and the model paints it back as a plausible face rather than that face. Your visual system is tuned for faces and catches it immediately, while an equally altered tree reads as fine.

Faces are the sharpest case of a wider rule. Anything that has to be exact at a small scale drifts first. Lettering on a sign, a familiar logo, or the pattern of a plaid shirt all live in the detail layer that noise erases, and all come back as approximations. Scenery is far more forgiving. That is why a landscape restyle works at a setting that would mangle a portrait.

Turning strength down further helps less than you would hope. Below roughly 0.2, the edit stops doing anything you can see. The fix is to take the face out of the edit entirely. Masked inpainting lets you regenerate everything except a protected region, so the untouched original pixels come back where it matters. We explain how that works in our guide to inpainting.

Practical recipes

• Restyle a sky: use strength 0.35–0.45. Write a prompt for the whole scene, sky included (“coastal town at dusk, dramatic orange storm clouds”). Describe the whole finished picture, including the parts you are keeping. The model repaints the entire frame, and your prompt is its only guide for every region.
• Photo to painting, composition intact: use strength 0.5–0.6, and name the medium first: “oil painting of a man reading by a window, visible brushstrokes, warm light.”
• Use a photo as a layout sketch: use strength 0.75–0.9 when all you want to keep is your rough arrangement of shapes. A quick snapshot of objects on a table becomes a still life in another world.
• Same settings, several seeds: at any strength, one starting number picks which plausible reinvention you get. (That number is the seed.) Generate a few and choose your favorite, rather than nudging the slider after every output.

The strength slider in Onhand

Onhand’s image-to-image editing on iPhone offers a 25%–85% strength slider for photo editing, labeled Subtle touch, Balanced, and Reimagine across the range. Different models interpret an edit differently, so treat the values above as starting points. Masked inpainting uses Turbo when you need to restrict the region. The source remains separate from the new result in your local gallery. Timing depends on the model and device; inspect important details after each edit.

PhoneDiffusion also does image-to-image on the device; Onhand vs PhoneDiffusion compares the two on editing tools, devices and price.