Inpainting, background work, retouching and upscaling with predictable results. Below are 7 copy-ready prompts. Fill in the [BRACKETS], copy, and paste into ChatGPT, Claude, Gemini or any capable assistant.
Editing prompts fail differently from generation prompts: the model has to preserve most of the image, and what it preserves depends on how you describe the change.
The 7 prompts
Plan an edit before touching the photo
Decide what a photograph needs rather than randomly adjusting sliders.
Help me plan an edit for this photo. THE PHOTO: [DESCRIBE IT - subject, lighting, what was captured, what conditions] WHAT IS WRONG WITH IT: [YOUR ASSESSMENT] WHAT IT IS FOR: [print / web / product listing / social / archive] SOFTWARE: [Lightroom / Photoshop / Capture One / phone app / AI tool] FILE TYPE: [RAW / JPEG / HEIC] MY SKILL LEVEL: [beginner / comfortable / experienced] Produce: 1. WHAT THE PHOTO IS ABOUT - the subject and the reason it was taken. Every edit decision should serve this. Editing without this leads to technically corrected photographs that are less interesting than the original. 2. THE PROBLEM LIST - ranked by how much it hurts the image. Separate: - FIXABLE: exposure, white balance, straightening, dust, distracting elements at the edge - PARTIALLY FIXABLE: soft focus, high noise, blown highlights (gone if truly clipped), heavy shadows - NOT FIXABLE: missed focus on the subject, motion blur on the subject, a fundamentally weak composition. Say so plainly rather than suggesting a fix that will not work. 3. THE EDIT ORDER - the sequence matters. Broadly: file handling and lens corrections, then crop and straighten, then white balance, then global exposure and contrast, then local adjustments, then colour, then detail and noise, then output sharpening last and sized for the destination. 4. THE SPECIFIC ADJUSTMENTS - for each problem, what to change and roughly by how much, in my software's terminology. 5. THE RAW ADVANTAGE - if I have a RAW file, what that allows that a JPEG does not: white balance without penalty, highlight and shadow recovery, and heavier adjustment before quality degrades. If I only have a JPEG, say what to be careful of. 6. THE CROP DECISION - whether to crop, what it gains, and what aspect ratio suits the destination. Crop is the most powerful edit and the most under-used. 7. WHEN TO STOP - the signs of over-editing: halos around edges, unnatural skin, crushed blacks, oversaturated colour, and obvious sharpening artefacts. Say what to watch for with these specific adjustments. 8. THE DESTINATION OUTPUT - export settings for my stated use: colour space, resolution, sharpening amount, and file size. 9. THE NON-DESTRUCTIVE RULE - work on a copy or use adjustment layers or a catalogue. Never overwrite the original.
What you get: A ranked problem list separating fixable from not, an edit order, specific adjustments and signs of over-editing.
Tip: Point 2's 'not fixable' category saves hours. No amount of sharpening recovers a photo where the focus missed, and knowing that immediately is worth more than any technique.
Write prompts for AI photo editing tools
Describe an edit to a generative editing tool clearly.
Help me write an editing prompt. THE IMAGE: [DESCRIBE WHAT IS IN IT] WHAT I WANT CHANGED: [THE EDIT] WHAT MUST STAY EXACTLY THE SAME: [WHAT TO PRESERVE] TOOL: [generative fill / inpainting / outpainting / object removal / background replacement / other] WHAT I HAVE TRIED: [IF ANYTHING] Produce: 1. THE EDIT TYPE - name precisely what operation this is, because the prompt differs completely between them: - REMOVAL: taking something out and filling plausibly behind it - REPLACEMENT: substituting one element for another - ADDITION: putting something new in - EXTENSION: expanding the frame - TRANSFORMATION: changing a property of what is there 2. THE SELECTION GUIDANCE - for masked edits, where the mask should go and how generous. Too tight leaves edge artefacts; too loose changes things you wanted kept. For removals, include a little of the surrounding area so the fill has context. 3. THE PROMPT - describe what should be in the edited area after the edit, not the action. Tools fill regions; they do not follow instructions like 'remove the bin'. Write what belongs there instead: 'continuous brick pavement in the same light'. 4. THE CONTEXT MATCHING - the prompt should specify the properties that must match the rest of the image: lighting direction, colour temperature, perspective, depth of field, grain, and time of day. Mismatched lighting is what makes an edit obvious. 5. WHAT WILL GIVE IT AWAY - the specific tells for this edit: repeated texture, a lighting direction that does not match, sharpness that differs from the surrounding area, perspective errors, and reflections or shadows that were not updated. Shadows are the most commonly forgotten. 6. THE SHADOW AND REFLECTION CHECK - if removing or adding an object, its shadow and any reflection must also change. Most tools do not do this automatically. 7. THE ITERATION PLAN - generate several, choose, and re-edit the remaining flaws rather than expecting one pass to be right. 8. WHEN TO DO IT MANUALLY - some edits are faster with a clone tool and a steady hand than with repeated generation. Say if this is one. 9. THE HONESTY QUESTION - if this image documents something real (a product, a property, an event, a news photograph), say what edits would misrepresent it. Removing a distraction is different from removing evidence.
What you get: The edit typed correctly, mask guidance, a prompt describing the result rather than the action, and the tells that reveal an edit.
Tip: Point 3 is the key mechanical insight. Generative fill prompts describe what should be there afterwards, not what you want removed.
Batch process a set of photos consistently
Edit many images to the same standard efficiently.
Help me batch edit a set of photos. WHAT THE PHOTOS ARE: [SUBJECT AND CONTEXT] HOW MANY: [COUNT] HOW THEY VARY: [lighting, settings, conditions - were they shot together or separately?] WHAT THEY ARE FOR: [PURPOSE] SOFTWARE: [WHAT YOU USE] FILE TYPE: [RAW / JPEG] Produce: 1. THE CONSISTENCY QUESTION first - were these shot under the same conditions? If lighting varied between shots, a single preset applied to all will make them worse, not better. Say what can be batched and what needs individual attention. 2. THE GROUPING - sort the set into groups that can share settings: by lighting condition, by location, by time, by subject type. Batch within groups, not across the whole set. 3. THE BASE EDIT - develop it on one representative image from each group, not the best one. The best image needs the least work and produces a preset that under-corrects everything else. Pick a typical one. 4. THE BATCHABLE ADJUSTMENTS - what can safely be applied across a group: lens corrections, a colour profile, a starting contrast curve, noise reduction, and a broad white balance if conditions were constant. 5. THE NON-BATCHABLE - what must be per-image: crop and straightening, spot removal, local adjustments, and exposure where it varied. Be realistic about how much time this takes for [COUNT] images. 6. THE WORKFLOW - the efficient order: import and apply the base settings on import, cull first (removing rejects before editing saves the most time), then group, then batch, then per-image passes, then export. 7. THE CULL FIRST RULE - editing before culling is the largest waste of time in batch photography. Say what proportion typically gets rejected and on what criteria. 8. THE SYNC MECHANISM - how to apply settings across a selection in my software, and the specific setting to exclude from syncing (crop and spot removal, usually). 9. THE CHECK PASS - reviewing the set together at small size to spot inconsistencies before export. Images that look right individually often do not look right as a set. 10. THE EXPORT - naming convention, sizing and format for the destination, and a plan for keeping the originals.
What you get: Photos grouped by shooting condition, a base edit from a typical image, batchable versus per-image adjustments and a cull-first workflow.
Tip: Point 3 is the counterintuitive one. Building your preset on the best photo produces settings that under-correct every other photo in the set.
Restore an old or damaged photograph
Repair a scanned family photo or a degraded image.
Help me restore this photograph. THE PHOTO: [DESCRIBE IT - approximate age, subject, format] THE DAMAGE: [scratches / fading / tears / water damage / mould / colour shift / low resolution / blur] HOW I HAVE IT: [scan resolution, or phone photo of the print] SOFTWARE AVAILABLE: [WHAT YOU HAVE] WHAT MATTERS MOST: [the faces / the whole scene / a specific detail] INTENDED USE: [print / share with family / archive] Produce: 1. THE SCAN QUALITY QUESTION first. Restoration is limited by what you started with. If this is a phone photo of a print, say what a proper scan would recover and whether it is worth redoing. Scanning at higher resolution before starting is frequently the single highest-value step. 2. THE DAMAGE ASSESSMENT - what is recoverable and what is not. Faded detail that still exists in the emulsion can be recovered; detail that is physically gone cannot be restored, only invented. Be explicit about which is which, because that distinction matters for a family photograph. 3. THE ORDER OF WORK: - Work on a copy; never edit the only scan - Structural damage first: tears, missing areas, large marks - Then dust and scratches - Then tonal recovery: contrast, faded density - Then colour correction or colour cast removal - Then noise and grain - Sharpening last, and gently 4. THE SPECIFIC TECHNIQUES for my damage types and software. 5. THE FACE PRIORITY - if faces matter most, they get manual attention and everything else gets automated treatment. Automated restoration tools often smooth faces into someone slightly different, which is the failure that matters most in family photographs. 6. THE AI UPSCALING CAVEAT - AI enhancement invents detail that was never there. For a decorative image that is fine. For a photograph of a person, it can change their face. Use it cautiously on faces, compare carefully with the original, and say what to check. 7. THE COLOURISATION QUESTION - if considering it: the colours will be invented, not recovered. That is a legitimate choice, but keep the original and label the colourised version as an interpretation. 8. THE ARCHIVE RULE - keep the untouched scan permanently, alongside the restored version. The restoration reflects today's tools and today's judgement; the scan is the record. 9. WHEN TO STOP - an over-restored photo looks artificial and loses its character. Some grain, some imperfection and some age are part of what it is.
What you get: A scan quality check, recoverable versus invented detail distinguished, an ordered workflow, and cautions on AI enhancement of faces.
Tip: Point 6 matters more than it sounds. AI upscaling of an old portrait can produce a clearer photograph of someone who is not quite the person in it.
Prepare product photos that convert
Edit product images for a listing or shop.
Help me prepare product photos. THE PRODUCT: [WHAT IT IS, MATERIALS, COLOUR] HOW IT WAS SHOT: [setup, lighting, background] WHERE THEY WILL APPEAR: [own site / Amazon / Etsy / marketplace] HOW MANY IMAGES: [COUNT] MY CONCERN: [WHAT LOOKS WRONG] SOFTWARE: [WHAT YOU HAVE] Produce: 1. THE ACCURACY RULE first, because it governs everything else. Product photos must show the product as it actually is. Colour accuracy, proportions, texture and any visible flaws are what determine whether the customer is satisfied or files a return. Editing that makes a product look better than it is buys one sale and costs a refund, a review and the customer. 2. COLOUR ACCURACY - how to get the colour right rather than pleasing: shoot or include a neutral reference, set white balance from it, and check against the physical product on a calibrated screen if possible. Say what to do if I cannot calibrate. 3. THE EDIT LIST - what to correct: - White balance and colour accuracy - Background: cleaning it, or replacing with white where the platform requires it - Dust, lint and fingerprints on the product - Exposure, so shadow detail and texture are visible - Straightening and perspective - Consistent sizing and product placement across the set 4. WHAT NOT TO EDIT - visible flaws that exist in the actual product, texture that indicates material quality, colour adjusted to look more appealing, and scale distorted by lens or crop. 5. THE IMAGE SET - what each slot should show: main image on the required background, in use or in context, scale reference, detail and texture close-up, all angles including the back, what is in the box, and any size or dimension graphic. Say which my [COUNT] images should cover and what is missing. 6. THE PLATFORM REQUIREMENTS - for my stated platform: background rules, minimum and maximum dimensions, whether text or graphics are permitted on the main image, and file format. Getting these wrong means rejection or poor display. 7. THE SCALE PROBLEM - customers consistently misjudge size from photographs, and it is a leading cause of returns. What to include: a familiar object, a hand, or a dimension overlay on a secondary image. 8. THE CONSISTENCY ACROSS THE SET - same white balance, same background, same product position and size, so the set looks deliberate rather than assembled. 9. THE EXPORT - dimensions, format and compression for the platform, balancing quality against page speed.
What you get: An accuracy-first edit plan with colour correction method, a what-not-to-edit list, the full image set and platform requirements.
Tip: Point 7 is the cheapest return-reducer available. Customers misjudge size constantly, and one image with a scale reference prevents a meaningful share of returns.
Remove or replace a background cleanly
Cut out a subject without the edges giving it away.
Help me remove or replace this background. THE SUBJECT: [WHAT IT IS - person, product, animal, object] THE EDGES: [hard and clean / hair / fur / fabric / transparent or reflective / motion blur] CURRENT BACKGROUND: [DESCRIBE IT - and how much it contrasts with the subject] NEW BACKGROUND: [what it will be placed on] TOOL: [WHAT YOU HAVE] OUTPUT: [where the final image goes] Produce: 1. THE DIFFICULTY ASSESSMENT - how hard this cut-out will be, based on my edge description. Hard-edged objects against a contrasting background are easy; hair against a busy background of similar tone is genuinely difficult and automated tools will not do it well. Set expectations honestly. 2. THE METHOD - for my subject and tool: automated selection, manual path, colour range selection, channel-based extraction for hair, or a combination. Say which to start with and what to fall back to. 3. THE EDGE REFINEMENT - the specific settings that matter: feather, contrast, shift edge, and decontaminate colours. Explain what each does, because they are the difference between a convincing cut-out and an obvious one. 4. THE COLOUR FRINGE PROBLEM - the old background's colour clings to the subject's edge, especially with hair and semi-transparent areas. This is the single most common tell. Say how to remove it in my tool. 5. THE MATCHING PROBLEM - a technically perfect cut-out still looks wrong on a new background if these do not match: - Lighting direction on the subject versus the new scene - Colour temperature - Contrast and black level - Perspective and horizon line - Depth of field: a sharp subject on a blurred background at the wrong scale looks pasted - Grain and noise level Address each. 6. THE SHADOW - a subject with no shadow floats. What kind of shadow the new scene requires: contact shadow where it meets a surface, cast shadow in the light's direction, or ambient occlusion. This is what sells the composite. 7. THE TRANSPARENT AND REFLECTIVE CASE - glass, water and shiny surfaces cannot be cut out cleanly because the background shows through them. Say what to do instead: reshoot on the intended background, or composite with the transparency preserved. 8. THE FILE FORMAT - keep transparency with PNG or a layered file, not JPEG. State where the edges will be destroyed by compression. 9. THE CHECK - view the result at 100% along every edge, and view it small to see whether it reads as one image.
What you get: A difficulty assessment, a method with fallbacks, edge refinement settings, the colour fringe fix, matching requirements and shadow guidance.
Tip: Point 6 is what separates a cut-out from a composite. A perfect edge with no shadow still reads as pasted, and a rough edge with a good contact shadow often does not.
Build a repeatable editing style
Develop a consistent look across your photography.
Help me develop a consistent editing style. WHAT I PHOTOGRAPH: [SUBJECT AND CONTEXT] WHERE IT IS SEEN: [portfolio / social / client work / product] THE LOOK I WANT: [DESCRIBE IT, OR DESCRIBE PHOTOGRAPHS YOU ADMIRE AND WHY] MY CURRENT EDITS: [DESCRIBE WHAT YOU DO NOW] SOFTWARE: [WHAT YOU USE] SHOOTING CONDITIONS: [consistent studio / varied natural light / mixed] Produce: 1. THE STYLE DECOMPOSED - turn my description into specific, adjustable properties: - Tonal curve: where the blacks sit, whether shadows are lifted, how highlights roll off - Contrast: global and local - Colour: saturation overall and per colour, hue shifts, and the colour of shadows and highlights - White balance tendency: warmer or cooler than neutral - Grain and texture - Sharpening approach A style is these choices, made consistently. Naming it does not help; specifying it does. 2. THE SIGNATURE ELEMENT - most recognisable styles have one dominant characteristic and the rest is restraint. Identify the one thing that should define mine, rather than applying five effects at once. 3. THE STARTING PRESET - the settings to build, with values, as a base to apply on import. 4. THE ADAPTATION RULES - a preset cannot be applied unchanged to photos shot in different light. For my stated conditions, what needs adjusting per image and what stays fixed. If my conditions are varied, be explicit that a preset is a starting point, not the edit. 5. WHAT THE STYLE SHOULD NOT DO - the limits. A style that fights the subject is a problem: heavy grading on product photography destroys colour accuracy, heavy contrast on portraits loses skin detail. Say where mine must yield to the subject. 6. THE CONSISTENCY TEST - assemble a set of edited images and view them together at small size. Do they read as one body of work? Where they do not, what is varying. 7. THE OVER-STYLING WARNING - the signs that the look has become the subject: identical colour grading regardless of scene, skies that are all the same, and skin tones that have drifted from plausible. 8. THE EVOLUTION - how to change a style over time without your back catalogue looking inconsistent, and whether to re-edit old work. 9. THE DOCUMENTATION - record the settings and the reasoning, so the style is reproducible in a year and after a software update.
What you get: A style broken into adjustable properties, one signature element, a base preset, adaptation rules and an over-styling warning.
Tip: Point 2 is what makes a style recognisable rather than busy. One strong characteristic and restraint everywhere else beats five effects stacked together.
Where AI actually helps here
- Background replacement and removal
- Removing objects from a scene with plausible fill
- Extending a frame for a different aspect ratio
Where it falls down
- Faces, which shift subtly between versions in ways people notice on someone they know
- Fine detail — text, logos, jewellery, patterned fabric
- Being told what not to change. Specify what should stay, not what should go
The mistake almost everyone makes: Describing the change instead of the result
‘Remove the car’ invites it to fill the space with anything. ‘A continuous brick wall where the car was, same mortar colour, same lighting’ tells it what should be there. Describe the finished image, not the edit.
Free tool: Image Prompt Builder
Runs in your browser. No sign-up, nothing uploaded.
Questions people ask
Why does inpainting change the whole image?
Usually the mask is too generous or the prompt describes the whole scene rather than the patch. Mask tightly and prompt only for what belongs inside the mask.
Can AI upscale a low-resolution photo?
It invents plausible detail rather than recovering real detail. Fine for a background; not for anything where the detail is evidence — faces, text, serial numbers.