How to edit a photo with AI instead of generating a new one
When to upload an existing image rather than start from a blank prompt, and how the brush works.
SamImageGenerator ·
The fastest way to get the image you want is often not to generate one. If you already have a photo that is ninety per cent right, editing it keeps that ninety per cent and changes the rest — where a new generation throws all of it away and starts again from noise.
Two different questions
Generating answers "make me a picture of this". Editing answers "keep this picture, but change one thing". The tell is in your own sentence: if it contains the words "this photo", you want the editing path. Upload the image and the request becomes an instruction rather than a description.
The brush is the whole feature
Paint over the part you want changed and write what should happen there. The instruction sent to the model becomes "change only the marked area", so the building under a replaced sky, the face beside a removed object and the light across the rest of the frame all stay exactly as they were.
| Goal | Mark | Write |
|---|---|---|
| Replace the sky | The sky | a dramatic sunset sky |
| Remove an object | The object and a small margin | remove, fill with the wall behind |
| Change a colour | The item | make the jacket dark green |
| Add something | The empty space | a potted plant in the corner |
Several areas can be marked in one pass, each with its own note, which is how "new sky here, remove the sign there" becomes a single credit rather than two rounds.
When to generate instead
If more than about half the frame is changing, or if the subject itself does not exist in any photo you have, generation is the shorter path. And if what you want is the same content in a different look, that is a restyle: pick a style with the photo uploaded and it becomes a look instruction rather than a new picture. The full set of editing jobs is on the edit styles page.