AI photo editing prompts that actually work

Four rules that make any prompt better

Name what to keep, not just what to change. These models re-render the whole image, so the only defence against unwanted changes is an explicit list of things that must stay. Faces, clothing, grain, lighting direction and text are the five worth naming.

Point at the area. Use the select or paintbrush tool where the app has one, since positional descriptions such as "the one on the left" are still read unreliably. Marking an area also limits how far the redraw spreads.

One change per message. Sequential single edits in one conversation hold together far better than one message requesting four things.

Say no as well as yes. Explicit prohibitions work better than vague requests for restraint. The words that change results most are below.

Clean-up and removal prompts

The most-used category, and the one where naming the fill material matters most.

• Remove the person walking in the background on the right. Rebuild the wall and pavement behind them. Do not change the subject, her clothing or the lighting. Names the target, the fill and the no-go zone in one line.
• Remove the bin and the cables on the left. Keep the same camera angle, the same depth of field and the same grain. The grain clause stops the patch looking smoother than the rest.
• Remove the power lines and the aerial from the sky. Keep the cloud shapes and the gradient unchanged. Sky is where generative fill is most reliable and least noticeable.
• Take the bottle off the table. Fill with the same tablecloth pattern and match the shadow direction already in the shot. Matching shadows prevents the floating clean patch.
• Erase the phone and my hands from the mirror reflection. Keep the mirror frame and the rest of the room reflection intact. The hardest removal to phrase, and the most requested.
• Remove every sign, logo and printed text from this image. Leave the architecture, materials and lighting exactly as they are. Useful for interiors and property, where branding dates a photo.

Light and colour prompts

Relighting is where generative models are genuinely convincing, because they are inventing something soft rather than something precise.

• Change the light to late afternoon sun from the left, long soft shadows, warm highlights. Keep every person and object in the same position. The position clause is what stops a re-composition.
• Neutralise the orange cast from the indoor lighting so skin reads as skin. Do not change the colour of the walls or the clothing. Fixes the most common colour problem in bar and restaurant photos.
• Make this look shot in overcast daylight: soft even light, no hard shadows, cooler colour, slightly lower contrast. The fastest way to match one odd photo to a set.
• Apply the colour grade of the second image to the first. Keep the first image's content completely unchanged. Two-image grading is the strongest use of multi-reference models.
• Lift the shadows until the faces are readable. Keep the highlights where they are and keep the mood dark. Stops it brightening everything into flatness.
• Give this a warm golden hour look with a gentle haze, without shifting skin tone or making the whites yellow. The exclusion is the half that does the work.

Portrait and skin prompts

The category with the highest failure rate. Short prompts with explicit preservation clauses do far better than detailed ones.

• Even out the lighting on her face. Keep freckles, pores, fine lines and flyaway hairs exactly as they are. Naming what to keep is the only reliable defence against plastic skin.
• Open the eyes in this frame, using the second photo of the same person as the reference for her eyes. The honest way to fix a blink, rather than letting it invent eyes.
• Reduce the shine on the forehead and nose. Do not change the shape of anything. Deliberately short, because narrow requests produce narrow changes.
• Remove the stray hair across her cheek. Leave the rest of the hair, including the flyaways at the crown, untouched. Small and verifiable, the kind of edit that survives inspection.
• Brighten the whites of the eyes and the teeth very slightly, subtle enough that it would not be noticed. Asking for restraint explicitly beats asking twice.
• Keep her face identical and change only the background. Do not adjust her jaw, nose, skin or hair. Worth pasting at the end of any portrait edit.

Background, product and reframing prompts

Where these models genuinely beat a conventional editor, because the job requires inventing something that was never in front of the lens.

• Place this product on a pale marble surface with soft window light from the left and a shallow depth of field. Keep the label text exactly as it is. Label text is where most product generations fall apart.
• Extend this image to 16:9 by generating more sky and sea on both sides. Do not add any objects, boats or people. The exclusion prevents a boat appearing on your horizon.
• Replace the cluttered background with a plain warm grey wall, matching the existing light direction on the subject. Matching the light is what makes a composite hold together.
• Reframe this to 4:5, keeping her fully in frame and generating whatever is needed above her head. Native aspect handling beats cropping and losing the composition.
• Put the subject from image one into the scene from image two, matching perspective, light direction and colour temperature. Those three matches separate a composite from a sticker.
• Generate three variations of this scene with the same product and the same lighting, different props each time. One batch request keeps a consistency that three separate prompts lose.

Restoration, text and document prompts

The highest-risk category, because this is where invented detail does real damage. Every prompt here contains a clause limiting invention, and you should keep the original scan regardless.

• Repair the scratches, creases and torn corner on this scanned print. Do not alter the faces or add detail that is not visible. The second sentence is the difference between restoration and fiction.
• Colourise this black and white photo with muted, period-appropriate colours. Keep the grain and do not smooth the faces. Muted beats vivid on every old photograph.
• Remove the yellow cast and the mould spots from this old print. Keep the original contrast and grain. Targeted damage removal rather than a blanket enhance.
• Add the headline SUMMER SALE in a bold sans-serif across the top third. Keep the photograph underneath unchanged. Short text in a named position is where text rendering is dependable.
• Translate the text on this menu into English, keeping the layout, typeface and background exactly as they are. One of the few genuinely reliable uses of a generative model on a real photograph.
• Describe everything visible in this photo in one plain sentence, for use as alt text. Not an edit at all, and the most useful thing a vision model does with a picture.

One prompt not to use, in any wording: asking a model to sharpen a document, a number plate or a face so you can read it. It will produce something legible that is not what was photographed.

A note on our own app: ShutterAI takes the same idea further with Recipes, saved looks you can share and reapply rather than re-describing a grade each time. It reaches iPhone this autumn, free to start, via the homepage.