Choose the repair before choosing an AI tool

First decide whether you need a faithful family record or a creative interpretation. For a family record, facial shape, expression, age, and visible personal details take priority over sharpness. Keep colorization or an imagined reconstruction in a separately labelled file.

What you see in the scanStart withLimit to respect
Faded tones, with facial detail still visibleLevels or curves in an image editorStop before dark hair or pale skin loses visible detail.
Dust or thin scratches on an otherwise intact areaSmall healing or clone repairs on a copyKeep the repair brush away from undamaged facial features.
Damaged background or an empty paper cornerA local repair, with AI as an optional source of replacement textureKeep the original faces and composition outside the repair.
A crease crossing a faceSeparate repairs along intact skin and feature edgesDo not let one broad edit redraw the eye, nose, and mouth together.
A missing eye, mouth, or large part of a faceA search for another print or a credible reference; otherwise retain the gapA plausible generated feature is not recovered evidence.
A tiny or blurred faceA better capture or a less compressed sourceMore output pixels do not establish the original facial detail.
A group photoOne person and one damaged area at a timeReview the smallest faces as carefully as the main subject.
Faded color, sepia, or a textured printA separate tone or color correction, then local cleanupDo not confuse the original photographic treatment with damage.

A layered editor such as GIMP gives you direct control over which pixels appear in the final image. ChatGPT can make an edited copy from an uploaded photo, but you still need to inspect it. For a photo whose faces must stay untouched, use the layered editor to assemble the final copy. The AI output can supply a background patch without becoming the whole photograph.

The GIMP masking steps below require a computer. If you are working only on a phone, capture the print carefully, request a small repair in your image app, and save the downloaded image separately. Zoom in on each face before keeping it. If the app changes a face, return to the unedited capture; for that photograph, either keep the remaining damage or continue in a desktop editor with layers.

Keep a clear, unedited scan

Work on a digital copy. Handle the print by its edges or housing on a clean, dry surface, and avoid touching the image. Do not tape a tear, scrub a stain, or force a photograph away from glass or an album page to make it easier to scan. The Library of Congress photograph-care guidance explains safe handling; its photograph FAQ advises consulting a photograph conservator for stuck photographs. Stop home handling if the surface is flaking, moldy, or otherwise unstable.

For a stable, flat print that fits comfortably, use a flatbed scanner. An album that cannot open safely needs supported copy photography. The National Archives digitizing guide warns against crushing an oversized original under a scanner lid and recommends a copy stand with supports for bound material.

If you have only a phone, photograph a stable loose print with the camera supported and parallel to it. Use even, diffuse light and keep reflections away from faces. Turn off flash, portrait effects, and beauty filters; avoid digital zoom. Keep all four edges in the frame, focus on the photograph, and compare more than one capture at full size. Move the light or camera to address glare instead of pressing or bending the print. Keep the camera's original file before applying a scan app's crop or perspective correction.

Capture the complete image and border. If you can turn the photograph over safely, capture the back separately, including names, studio stamps, and dates. Keep the unedited capture even if a display version later removes the border. Disable automatic face enhancement or restoration for this master so you have a record of the source before those edits.

Check the capture at 100% before putting the print away. Can you see detail that is visible on the physical photograph? Are there reflections, clipped edges, or uneven lighting? Recapture those problems while the original is available. Raising a scan's resolution after capture does not rescan the print.

For ordinary snapshot prints, the Library of Congress scanning introduction gives 300 to 400 dpi as suitable for many personal uses. Small faces or planned enlargements may justify a higher-resolution capture. For example, a 2 × 3 inch print scanned at 600 pixels per inch produces approximately 1200 × 1800 pixels. Inspect the result rather than assuming a higher number recovered more facial detail. Negatives and slides need a capture method suited to film.

Choose a photo or continuous-tone scanning mode. Use grayscale for a neutral black-and-white print, or color when you want to retain sepia, paper tone, or an existing color image. Avoid a document setting that reduces everything to solid black or white; a face needs the shades between them.

Keep this unedited scan as your master copy. Save it as a lossless TIFF or PNG when your capture software supports it. If the only source is a JPEG, keep that original file too. Converting it to PNG can prevent another lossy save during editing, but cannot undo its existing compression. Give working files separate names, such as family-portrait-scan-master.tif and family-portrait-repair-v01.xcf.

Write down the details a repair must preserve

Before editing, inspect each face and note the details you can actually see. Useful notes are specific: the left eye is slightly narrower, the mouth is closed, the hair parts on the subject's left, the glasses have thin round rims. Include a mole, scar, or uneven tooth only when the source or reliable family information establishes it. Do not ask the model to complete an uncertain feature from your description.

Also check jewelry, clothing fasteners, hands, and the position of each person. A small pin can matter to the family, and changing a glasses frame can alter both the face and the apparent period. Distinguish a scratch from a wrinkle before removing it. If you cannot tell, mark the area as uncertain and leave it for review.

For a group, identify people by position rather than an unconfirmed name. A short note can be enough:

Left person: closed mouth; hair part and round collar pin stay unchanged.
Center person: preserve brow shape, hairline, and visible mark above eyebrow.
Right person: preserve thin glasses, mouth shape, necklace, and apparent age.
Uncertain: crease crosses part of the center person's face. Do not invent the obscured detail.
Allowed: background scratches, dust, and a small blank paper corner.

Keep this note beside the scan while you work. An AI description can help you organize observations, but it should not decide whether an unclear mark is damage or part of the person's appearance.

Try tone correction and small repairs before generating a new image

When the face is present but hard to see, start by revealing the information already in the scan. In GIMP, open a working copy, duplicate the image layer, and keep the lower source layer unchanged. Save the layered project so you can return to individual edits.

For fading, use Colors → Levels. The GIMP Levels documentation describes the black, white, and midtone controls. Adjust them in small increments while watching facial shadows and highlights. Moving the black point too far can merge eyebrows or dark hair into a solid patch; pushing the white point too far can erase pale skin detail. Use the preview to compare with the unadjusted image. There is no single Levels setting that suits every old print.

For a speck or thin scratch surrounded by intact texture, use the Heal tool. Choose a brush only slightly wider than the damaged spot, Ctrl-click a nearby clean source, then make a short repair. Sample again as the surrounding texture changes. A long stroke across skin, hair, and clothing borrows from areas that need different textures.

The Clone tool copies source pixels more directly. It can help where healing smears a sharp boundary, but the source must match the destination. Copying the other eye or repeating a section of a necklace can create a convincing error. Undo a repair that changes a feature's outline, even if it removes the scratch.

Check each repaired area at 100%, then look at the whole portrait. A patch can blend at high zoom yet look like a flat spot at normal size. If tone correction and local cleanup solve the problem, you can finish without generating a replacement photograph.

Ask AI for a small, specific repair

Before uploading a private family photo, check whether you are comfortable with the service's current upload, retention, and sharing settings. Obtain appropriate permission for recognizable living people. Keep sensitive inscriptions and metadata in your private master; use a separate upload copy if those details are unnecessary for the repair. A local editor is a useful option when you prefer to keep the photograph on your computer.

In ChatGPT, attach the working scan and ask for a specific edit. OpenAI's image-generation guidance describes editing from a reference image and making targeted revisions. Request one edited image so you can inspect it before trying further variations. Describe the damage and its location instead of asking for a beautiful, modern, or high-definition portrait.

Conservative restoration prompt
Edit the attached photograph as a conservative digital restoration.
Repair the visible background scratches, dust spots, and damaged blank paper at the lower-left corner. Make only a restrained correction to fading.
Preserve every person's facial proportions, expression, gaze, apparent age, hairline, hairstyle, glasses, jewelry, and clothing. Keep the original framing and arrangement of people.
Keep the source's monochrome or sepia treatment. Do not add a smile, teeth, eyelashes, makeup, smoother skin, or new facial detail. Leave unclear facial areas unresolved.
Return one edited image.

Change the locations and allowed repairs to match your photograph. Keep the instructions about uncertain faces. Do not include a fictional accessory from this example in a prompt for your own family.

Save the result under a new filename, then inspect the downloaded file's dimensions. It may be smaller than the scan. Keep both files, and leave the original scan at its captured size when you bring the edit into a layered project.

To retry one area, select the damage if the editor offers that control, and describe only the repair needed there. Inspect the rest of the frame afterward. OpenAI's GPT Image mask documentation says the model treats a mask as guidance and may not follow its exact shape. Check faces and objects outside the selection too.

Repair only the vertical scratch in the plain background to the right of the seated person. Match the surrounding texture and tone. Leave the person, glasses, hair, clothing, and all other areas unchanged.

If a retry changes a face or loses an accessory, return to the original scan or the last good layered version. Editing the altered face again makes it harder to tell which details came from the photograph.

Keep original faces while using an AI background repair

When you need unchanged facial pixels, assemble the final copy in a layered editor. Keep the original scan as the bottom layer and place the AI edit above it. GIMP's layer masks determine which parts of the upper layer are visible. Black on the mask hides the edit; white reveals it.

  1. Open the master scan in GIMP. Import the AI edit with File → Open as Layers. Rename the lower layer Original scan and the upper layer AI repair.
  2. If both images have the same crop and proportions but different pixel dimensions, select AI repair and use Layer → Scale Layer. Keep the width-height chain linked. For the example files, setting its width to 1600 pixels also gives a height of 1200. Resize only the AI layer.
  3. Temporarily lower that layer's opacity to inspect alignment against the original. Background edges should line up. If the crop or shapes have changed and simple scaling cannot align the repair, discard that patch. Restore the layer's opacity to 100% before continuing.
  4. Select AI repair and choose Layer → Mask → Add Layer Masks. Choose black, full transparency. The AI edit disappears and you see the original scan.
  5. Click the new mask thumbnail, choose the Paintbrush, and set its foreground color to white. Paint only over the background scratch or blank corner you want to replace. Leave faces, ears, hairlines, and glasses black on the mask.
  6. Inspect the patch boundary at 100%. If you reveal too much, paint black on the same mask to hide that part again. Keep any soft transition at a background seam away from the face.
  7. Use the eye icon beside AI repair to switch it off and on. The repaired background should change; each face should continue to come from Original scan. Save the layered project before exporting a copy.

This method leaves damage inside the protected face in place. Repairing that damage requires a separate decision about which facial pixels you are willing to alter. If a crease has erased an eyelid, keeping the face faithful may mean accepting a visible crease.

If you see white paint on the photograph, undo it and click the mask thumbnail instead of the image thumbnail. If no repair appears, check that the brush color is white and the AI layer is visible. GIMP's mask-editing control determines whether your brush changes the mask or the image.

If you require identical facial pixel values, keep the original layer unfiltered and untransformed, leave the AI layer fully hidden over every face, and export without resizing or color conversion. Save a lossless copy. JPEG compression can change pixel values even where you made no visible edit. For a family album, the practical review is whether the facial shapes, expression, and age still match; you can make gentle tonal changes while keeping those features intact.

Handle the difficult cases separately

A crease crosses a cheek, an eye, or the mouth

Break the crease into sections. Plain cheek texture offers nearby material to sample; an eyelid or lip edge requires its own shape to remain visible in the source. Repair a short section, compare it, then continue. A single broad face repair can change several features at once.

If the crease hides the corner of the mouth, do not allow the repair to decide whether the person was smiling. Look for another copy of the same photograph. Another portrait can help a reviewer recognize the person, but differences in age, pose, and expression mean it cannot establish the exact missing feature in this image. Record any reconstruction separately.

You have another photograph of the same person

Another print of the same exposure is the most useful reference for missing detail. Check that it really is the same photograph: similar portraits may have been taken moments apart with different expressions. If it is a different portrait, use it to check recognizable features and discuss the damage with a restorer.

If you upload two images for an AI edit, assign their roles explicitly. For example: “Edit image A, the damaged group photo. Image B is a separate portrait for comparison only. Keep image A's pose, expression, age, and lighting; do not paste image B's face into it.” Rebuilding a missing eye from another pose still involves interpretation. Save that version separately from the copy that preserves the visible evidence.

The face is blurred or very small

Look for the original scan, negative, or a better file before using face enhancement. A messaging-app copy or screenshot may contain less information than the file held by another relative. If you still have the print, check whether a better capture reveals detail.

Compare an enhanced face with the source at the same display size. Added eyelashes, iris texture, teeth, or skin detail may look persuasive while remaining unsupported. If the mouth shape or eye spacing changes, reject that result. A softly rendered face can be the more faithful version. For a family album, a smaller print may be preferable to a large enlargement filled with invented detail.

There are many people in the photograph

Number people from left to right and review every face. A good result on the front row does not establish that the small faces behind it survived. Check people near damaged edges, hats, veils, and overlapping shoulders. Compare the person count, gaze, and separation between adjacent faces.

Work on one damaged area at a time and preserve the full group scan. If you use a crop to inspect or edit a face, keep its connection to the original position. Avoid replacing a whole row of people because one face is damaged. Cropped repairs still need alignment and edge checks when returned to the group.

A corner, hand, or part of a body is missing

Blank paper and a plain backdrop contain fewer identity details than a missing hand or medal. Limit an AI fill to an area whose replacement you can accept as an approximation. If the gap includes a person, an inscription, or a distinctive object, leave it visible in the faithful copy unless you have adequate reference evidence.

You can make a separate cropped display version when the missing corner is distracting. Keep the complete scan and note the crop. Extending the frame to invent clothing or a room changes the scope from cleanup to reconstruction.

A color photo has faded, or a monochrome print is warm

Separate color correction from damage repair. For an existing color photograph, compare gentle channel adjustments with the source and watch skin, neutral objects, and clothing. Do not assume that a modern-looking skin tone establishes the original color.

A warm monochrome print may have an intentional photographic tone. Preserve that appearance in the master and save any neutral black-and-white conversion as a separate copy. For an originally monochrome photo, generated colors are interpretations unless independently documented. Keep a faithful monochrome version and label an optional colorized copy so a later relative does not mistake it for a color original.

The scan has paper texture, grain, or many white specks

Check whether the pattern covers the whole image consistently or occurs as isolated damage. Removing every small mark can flatten a textured print and erase fine facial details. Try a small background area first. If cleanup makes hair clump together or skin look waxy, reduce it or return to local spot repair.

Inspect boundaries around glasses and jewelry after denoising or sharpening. A bright rim, dark outline, or repeated patch of texture is a reason to revisit that edit. Judge the face at normal viewing size as well as at 100%; sharpening a damaged edge can make the damage more conspicuous.

No relative can confirm the likeness

Use the visible scan as your reference and keep the changes narrow. Compare proportions and expression without claiming that someone has confirmed the person's appearance. Ask reviewers to distinguish what they remember from what they infer from the image. When reviewers disagree about an uncertain feature, preserve the notes and the less reconstructed version.

The damage is too extensive for a small repair

For extensive damage in a digital scan, ask a photo restorer to assess one difficult area before committing to a full restoration. Explain that preserving the person's appearance matters more than making every area sharp. Ask which details can be repaired from the scan, which would be reconstructed, and whether you will receive both the edited file and the untouched scan.

Physical damage needs a different specialist. Stuck, flaking, or moldy photographs should be assessed by a photograph conservator before further handling. A digital retouch can make a copy easier to view, but cannot stabilize the original print. The Smithsonian's family-archives advice explains what to look for in a service provider, including experience, quality checks, and safe handling.

Review the faces before accepting the cleanup

Open the scan and edited copy at the same scale. Compare the whole frame first, then each face at 100%. If you use a layered comparison, make a temporary aligned review copy and toggle between the images. Do not overwrite or distort the original to make a comparison fit.

InspectReason to reject or undo the edit
Eyes and browsChanged eye spacing, gaze, eyelid shape, brow angle, or newly invented eyelashes
Nose and mouthA narrower nose, lifted mouth corners, new teeth, or altered lip shape
Age and skinLost wrinkles or marks, a changed jawline, or uniform plastic texture
Hair and glassesA different hairline or part, missing hair, altered frames, or changed lens boundaries
Clothing and objectsA replaced brooch, new button pattern, missing necklace, or invented lettering
Whole frameA missing person, shifted pose, changed crop, repeated texture, or edits outside the intended repair

Ask a knowledgeable relative specific questions: Does this person still have the same expression? Does the repair make them look younger? Is that mark part of the face? Show the scan beside the edit without introducing one as the better version. Record objections per person so a disagreement about one face does not disappear into a general approval of the photograph.

Accept only repairs that survive those comparisons. If the background improves but a face changes, keep the background patch through masking and retain the source face. If the edit offers no clear benefit without new uncertainty, keep the simpler version.

Compare a restrained repair with an over-edited portrait

These images use a fictional AI-generated portrait with simulated damage. No real family photo was used. Compare the expressions, the younger woman's collar pin, and the older woman's glasses.

Fictional monochrome family portrait with simulated scratches, fading, dust, and a missing paper corner
Damaged scan. The damage and people are fictional. Credit: Grow AI Skills. Open full-size image.
Over-edited version of the fictional portrait with added smiles, color, younger faces, and changed accessories
Over-edited version. Compare the smiles, collar pin, and glasses. Credit: Grow AI Skills. Open full-size image.
Restrained repair of the fictional family portrait in muted monochrome, shown for visual comparison
Restrained repair. Compare each face with the scan before keeping it. Credit: Grow AI Skills. Open full-size image.

In the over-edited version, the expressions change, the people appear younger, the scene gains color, and the brooch and glasses change. The restrained repair avoids those obvious transformations. It still needs a close comparison with the scan; looking more plausible does not establish that every facial detail is correct.

The damaged scan is 1600 × 1200 pixels. Both AI edits are 1448 × 1086 pixels, so neither keeps the scan's full pixel dimensions. Open the individual images at full size before comparing small details.

To practice, download the damaged scan and try a background-only repair while retaining the original faces. You can use the family-review worksheet to record a separate opinion about each person.

Example images: Grow AI Skills. The editing guidance combines official documentation with practical recommendations; this guide does not report a real family-photo restoration test.

Export for the album, the screen, or a print

Keep the unedited scan, the layered working project, and the selected restoration as separate files. Add a brief note identifying the source, edits, uncertain areas, reviewer, and review date. If you made a colorized or reconstructed version, give it an explicit filename such as family-portrait-colorized-interpretation.jpg.

In GIMP, use File → Save As to retain the editable .xcf project with its layers and masks. Use File → Export As for a separate PNG, TIFF, or JPEG. Reopen the exported file and check its dimensions and appearance before sharing it.

For sharing, export a JPEG copy from the selected version and keep the lossless file. Avoid repeatedly editing and saving the same JPEG. If you remove private metadata for sharing, do so on the sharing copy while retaining the family record privately.

For printing, divide pixel dimensions by the intended pixels per inch. The example's 1448 × 1086 repair gives approximately 4.8 × 3.6 inches at 300 ppi. A 6 × 4 inch print at 300 ppi requires 1800 × 1200 pixels; filling that shape from a 4:3 image also requires a crop, so check that no head or meaningful edge gets cut off. A lower print density or a border is another option. Upscaling can provide a larger file, but does not verify new facial detail.

Inspect a small proof print before ordering an enlargement. Look for faces that become too dark, pale areas that lose detail, and repair seams that were hard to see on screen. Keep the physical original separately. For the digital files, the National Archives recommends a 3-2-1 backup arrangement: three copies, on two different media, with one copy off-site.