When you hand a model an existing image, it does not start from pure noise. It adds noise to your picture and then denoises back out. Strength is how much noise it adds: at a low value the original is still clearly visible under the static, so the model has little freedom; at a high value it is almost gone and the model is effectively generating fresh with a vague memory of the layout.
That gives the setting a very practical reading. Small values are corrections — grade shifts, texture cleanup, a slightly different expression. Middle values are restyles that keep the composition. High values are new pictures that happen to share a silhouette.
Identity is the thing that breaks first. Faces stop being the same person well before layout stops being the same layout, so if a specific person has to survive the edit, the workable band is much narrower than it looks.
In practice
- Bracket it: run the same edit at low, medium and high and pick, rather than guessing once.
- It interacts with step count — very few steps at high strength gives you noise the model never finished resolving.
- For a targeted change, masking the region beats raising the strength.
The mistake to avoid
Raising strength because the model is not making a change you asked for. If the prompt term is being ignored, more freedom just produces a different picture that also ignores it.
Related terms
Image-to-image
Image-to-image takes a picture as its primary input and returns a changed picture — restyled, corrected or varied — instead of inventing one from nothing.
Inpainting
Inpainting regenerates a region you have masked while leaving the rest of the picture untouched, so a change stays local.
Sampling steps
Sampling steps is how many passes a model takes to turn its starting noise into a finished output — more passes, more refinement, more time.
CFG scale
CFG scale controls how strictly a model obeys your prompt, trading obedience against the model's own sense of what a natural image looks like.
Video-to-video
Video-to-video takes finished footage in and returns altered footage, using the original clip as the structural reference for every frame.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.