AI restoration can make a damaged photograph easier to read, but it cannot know every missing eyelid, tooth, or wrinkle from one small source image. If preserving a person’s likeness matters, treat the result as an interpretation to review—not as proof that the original face has been recovered. The safest workflow keeps the scan, compares a light restoration with the source, and rejects any change that alters a distinctive feature.
AI-generated illustration. The scene is fictional and shows a checking process; it is not a Photo Editor AI result or a real before-and-after comparison.
What “without changing faces” can and cannot mean
The phrase is a quality goal, not a switch that can be proved from a single output. A restoration model may remove contrast loss, dust, and small breaks while also filling missing pixels with a plausible shape. That shape can look convincing and still be wrong. The older, smaller, more damaged, or more oblique the face, the less evidence the file contains for an exact reconstruction.
Start by making a preservation copy of the scan. Do not overwrite the original with the processed file. The Library of Congress recommends a clean workspace and careful handling by the edges or housing; those habits matter before any upload. If the physical print is brittle, stuck to glass, moldy, or historically valuable, stop and ask a photograph conservator rather than flattening or cleaning it yourself.
For a useful review, define what must stay recognizable. That may be the spacing of the eyes, the shape of the mouth, a hairline, a birthmark, a prominent nose, or the apparent age of the person. Do not use generic smoothness as the test. A sharper face is not necessarily a truer face.
A cautious restoration workflow
- Scan or photograph the original once, then archive that file. Keep the full frame, including edges and handwritten context, even if you later crop a working copy. Record which side is the front and any known date or identity separately from the image.
- Choose a single working image. Photo Editor AI’s Photo Restoration page is designed for one JPEG, PNG, or WebP at a time and shows a 10 MB upload limit for this feature. Keep a copy below that limit without replacing your archive file.
- Use the least ambitious composition that fits. The tool exposes
auto, square, portrait, landscape, and phone-shaped aspect ratios. Pick the ratio that avoids unnecessary cropping. Do not assume changing the ratio creates missing photographic evidence. - Save the result as a separate candidate. The page is configured to request a 1K JPG, but that request parameter does not verify the dimensions or quality of the file you receive. Check the downloaded file’s actual dimensions and do not expect a face-preservation control. Your review is the decision point.
- Compare at two scales. First compare the complete original and complete candidate at normal viewing size, asking whether the expression and age feel consistent. Then open the complete images at 100% and inspect the face closely: eyes, pupils, nostrils, teeth, hairline, ears, and any feature relatives use to identify the person. If a crop helps you see the face, make it only as a duplicate close-review crop; do not use it to replace the full-image comparison.
The Photo Restoration tool is a reasonable next step for a copy you are prepared to evaluate. Your own source quality and damage pattern will determine whether the candidate is useful, so keep the original beside it while you decide.
The likeness checklist: what to accept and what to reject
Use a simple three-column note: source evidence, candidate change, and decision. A candidate can pass one area and fail another. For example, a smoother background may be helpful while a newly shaped eye is unacceptable.
Check these features in order:
- Face outline: Has a shadow or crease been turned into a different jaw, cheek, or hairline?
- Eye geometry: Are the eyes still the same distance apart, with a believable lid and gaze? Artificially symmetrical eyes are a warning sign.
- Mouth and teeth: Has a dark gap become a new tooth or a closed smile? Missing teeth should remain uncertain.
- Age cues: Did the candidate erase every wrinkle or add an implausibly smooth face? Restoration should not silently change the person’s apparent age.
- Distinctive marks: Compare moles, scars, glasses, earrings, and hair parting. If a mark is absent in the source, do not treat a new mark as recovered history.
- Edges and hands: Models can make fingers, collars, and glasses melt into nearby texture. A face can appear plausible while the surrounding anatomy signals that the file needs rejection.
Reject the candidate when a central feature changes, when repeated texture forms an invented pattern, or when you cannot explain a detail from the source. Keep it only when the change is a modest cleanup and the identity checks remain consistent. When the face is mostly missing, label the result as an artistic reconstruction if you keep it at all.
Failure branches and safer alternatives
If the face is crossed by a deep crease, review the complete photo first, then make a duplicate close-up crop only for examining the face at 100%. A broad automatic repair may bridge both sides of the crease with a new eye or mouth. For a small mark on a plain background, a local healing tool can be easier to control; Adobe’s documentation describes Spot Healing and Healing Brush workflows that blend sampled nearby pixels. That is a manual alternative, not evidence that an AI restoration will make the same choice.
If the photograph is too dark to distinguish eyes or the scan is heavily blurred, do not ask a restoration tool to invent certainty. Find another family copy, a negative, a higher-quality scan, or a relative who can identify the person. If no supporting evidence exists, keep the source and describe the uncertainty in the file name or notes.
If colors are faded, handle color as a separate question. A believable skin tone does not prove the historical color was correct. Use the faded-color workflow after you have decided whether the face itself remains trustworthy. For scratches and folds, the damage-specific guide explains why location matters.
How to keep the history with the image
Keep the untouched scan, the candidate restoration, and a short review note together. A useful note says what was changed, what remains unknown, and which face features were checked. If the file will be shared with family, say when it was AI-assisted so a plausible invented detail is not repeated as a fact about the person.
The goal is not to make every old photo look new. It is to make a defensible choice about what the source supports. A modest cleanup that leaves the face recognizably consistent is often more valuable than a dramatic reconstruction that looks polished but changes who the viewer sees.