How to Edit a Photo With AI Without Losing the Person

Most AI photo edits fail for a boring reason: the instruction asked for too much at once. The face changed, the hands multiplied, the logo turned into gibberish, and the background no longer matched the light. The model did what it was told, but the prompt never told it what to protect.

This guide is a practical editing workflow. It assumes you already have a photo and a specific change in mind. It does not assume you want a new person, a new scene, or a completely different picture.

The short answer

Write one sentence that names what must stay the same, then one sentence that names the single change you want. Upload the sharpest version of the original. Compare the result with the original at full size before you publish it. If the face, hands, text, or edges drifted, run the edit again with a narrower instruction instead of accepting the first result.

That is the whole method. The rest of this guide explains why each part matters and how to apply it to real photos.

What an AI image editor actually changes

An editor that works from a prompt does not paint inside a selection the way a traditional retoucher does. It reads the image and the instruction, then generates a new image that should satisfy both. That means it can change anything, including parts you wanted untouched.

A useful mental model: the model is a very fast illustrator with a copy of your photo on the desk. It will try to keep the photo recognizable, but it is allowed to redraw whatever it thinks is necessary. Your prompt is the brief that tells it where the line is.

This is different from a filter. A filter applies a fixed transformation to every pixel. A prompt-based edit is closer to a directed redraw: it can remove an object, replace a background, change clothing, or relight a face, but it can also reinterpret details you did not mention.

Step 1: prepare the source photo

The source sets the ceiling for the result. A sharp, evenly lit photo with a clear subject gives the model useful information. A blurry, dark, heavily filtered photo gives it room to invent.

Use these checks before uploading:

• The subject is in focus, especially the eyes.
• The face is not heavily smoothed by a beauty filter.
• The file is large enough to inspect at 100 percent.
• There is no important text that must survive the edit.
• You have the rights to edit and publish the image.

If the file is too large, compress it first. If it is a screenshot with a tiny face, do not expect a portrait-quality edit. The model cannot recover detail that was never captured.

Step 2: write a preservation sentence

The preservation sentence is the part most people skip. It is also the part that prevents the most common failures.

A preservation sentence names the things that must not change:

Preserve the subject's recognizable face, facial proportions, skin tone, approximate age, body proportions, and natural anatomy. Keep the same number of people as the uploaded photo.

For products, the sentence changes:

Preserve the product's exact shape, color, material, label text, and logo placement.

For interiors:

Preserve the room geometry, window positions, floor lines, and the direction of the existing light.

The model does not need a legal contract. It needs a clear constraint. One or two sentences are enough.

Step 3: write one change sentence

Now describe the change. Be concrete about the visual result, not the tool operation.

Weak: "Make it better."

Weak: "Improve the background."

Useful: "Replace the background with a quiet studio wall in warm gray, add a soft contact shadow under the subject, and keep the existing light direction."

Useful: "Change the jacket to a dark navy blazer, keep the shirt and pose unchanged, and keep the face and hair identical."

Useful: "Remove the small bin and the cables near the left wall, continue the floor and wall texture naturally, and do not move the furniture."

One change per run is easier to judge. If you need three changes, run three edits in sequence, checking each one. The cost is higher, but the failure is easier to locate and repair.

Step 4: describe the light, not just the objects

Light is what makes an edit look pasted or believable. When you add a new background, say where the light comes from and how soft it is.

For a portrait:

Soft window light from the left, gentle falloff on the right side of the face, natural skin texture, no beauty retouching.

For a product:

Large softbox from the upper left, a subtle fill from the front, a soft contact shadow to the lower right, no harsh specular highlights on the label.

For an interior:

Warm late-afternoon light from the windows on the right, keep the shadows in the corners, do not add a second light source.

If the original light is strong and directional, a flat new background will look wrong. Match the direction first, then adjust the color.

Step 5: inspect the result like an editor

Do not judge the result from the thumbnail. Open it at full size and check the areas that fail most often.

If one area fails, do not rerun the whole edit with a longer prompt. Rerun it with a narrower scope. For example, if the background is good but the hand is wrong, ask for the same background and explicitly protect the hands, or crop the edit to the area that needs work.

A reusable prompt template

Here is a template that works for most portrait edits:

Use the uploaded photo as the primary reference.
Preserve the subject's recognizable face, facial proportions, skin tone,
approximate age, body proportions, pose, and natural anatomy.
Keep the same number of people as the uploaded photo.
[One specific change.]
[Lighting and surface details.]
Photorealistic editorial photography, realistic skin texture, natural shadows.
Avoid altered identity, extra limbs, duplicated fingers, distorted anatomy,
text, logos, and watermarks.

Replace the bracketed lines. Keep the preservation and avoidance lines unless you have a reason to change them.

When to use a different workflow

Prompt-based editing is not the right tool for every job.

• Old damaged print with scratches and missing corners: use a restoration workflow, and keep the original scan.
• Photo needs more canvas, not a different scene: use an extender or outpainting workflow.
• Photo is technically weak but the scene is fine: use an enhancer workflow with a narrow correction.
• You need a transparent cutout: use a background-removal tool, not a generative replacement.
• You need a precise measured change: use a traditional editor. A generative model is an interpretation, not a measurement.

Common mistakes

Asking for a new scene and a new face at the same time. The model has no reason to keep the identity if the prompt does not protect it.

Using "professional" as the only instruction. Professional to the model can mean a blazer, a gray backdrop, a different age, or heavy retouching. Name the visual result you want.

Editing a low-resolution screenshot. Small faces and compressed text are the hardest inputs. Start from the original file.

Accepting the first result because it looks good at thumbnail size. Most failures are visible only at full size.

Publishing an edited photo as a documentary record. An edit is an interpretation. If the image is used as evidence, label it and keep the original.

A five-minute checklist

Before you export, run this list:

1. Does the person still look like the person in the original?
2. Are the hands, ears, and teeth anatomically plausible?
3. Is any text in the image still correct?
4. Do the edges of the subject hold up against the new background?
5. Does the shadow direction match the light?
6. Is the result honest about what it is?
7. Do you still have the original file?

If any answer is no, the edit is not finished.

Try the workflow

The AI image editor page opens the workspace in image-to-image mode with three starter prompts. You can read the page and copy a prompt without an account. A generation uses credits, and the pricing page lists the current packs and plans.

If you need a wider frame instead of a changed scene, read the image extender guide. If the background is the only problem, the background changer guide is a narrower starting point.