FIELD NOTE
How to keep a recognizable identity in an AI portrait
A careful workflow for reference images, identity locks, group mapping, and corrections when a generated face drifts.
Choose a reference that gives the model evidence
A reference does not need to be professionally lit, but the face should be visible and unobstructed. A forward or three-quarter view with even light gives a model evidence about the eyes, nose, mouth, face shape, and skin texture. Very small faces, heavy filters, sunglasses, and motion blur remove the details that an identity instruction is trying to preserve.
Use one strong reference for a single-person portrait before adding a second image. Multiple references can help with different angles, but they can also create competing signals. If you need several references, say which image is the identity anchor and which image only supplies clothing, pose, or background.
Separate identity from style
Write the identity sentence first: preserve recognizable facial structure, approximate age, skin tone, and distinctive features. Follow it with the style change: an 1980s haircut, tailored jacket, or mall portrait backdrop. This ordering tells the model that the decade styling is a transformation applied to the person, not a new person who merely shares a mood.
Avoid absolute claims such as “make the face exactly identical.” They set an impossible target and can encourage artifacts. The useful promise is narrower: preserve the reference features while changing only the requested clothing, pose, light, and setting.
Map people in a group
Group prompts need a map. State the exact number of people, then assign positions from left to right or front to back. Name each reference as person A, person B, and so on, and repeat that mapping in the correction prompt. This reduces face swapping and prevents the model from adding an extra person to fill an imagined composition.
Keep the group action simple on the first pass. Ask for relaxed overlapping poses and all hands visible rather than a complex movement sequence. Once the people and positions are correct, use a second pass for a prop, a different crop, or a stronger period detail.
Diagnose drift instead of rewriting everything
When a face changes, identify the failure. If the eyes or mouth are wrong, re-anchor the face and shorten the style list. If the person is correct but the image looks modern, keep the identity instruction and add film, flash, wardrobe, and object corrections. If two people have blended features, restate the left-to-right mapping and exact count.
A focused correction is easier to evaluate than a completely new prompt. Save the original output, the correction text, and the new output together. This makes your own workflow repeatable and helps identify which models need a closer crop or a different reference.
Respect permission and context
A recognizable face is personal data in many contexts. Use your own images or obtain clear permission before transforming another person's likeness, especially for public, commercial, political, or sensitive uses. Do not present a generated portrait as a real event or as an endorsement by the person shown.
The site provides identity guidance for creative work; it does not grant rights to a reference image. Read the model provider's rules and the laws that apply to your use. When in doubt, choose a fictional subject or a reference you control.
Know when to stop
Some combinations will remain unstable. Extreme profile views, tiny faces, old low-resolution scans, and large groups can exceed a model's reliable range. A good workflow has a stopping rule: if three focused corrections do not improve identity, change the reference or simplify the scene instead of adding more adjectives.
Consistency is a practical goal, not a promise of perfect duplication. The strongest result usually comes from a clear reference, a restrained scene, and one controlled change at a time.
Ready to try a prompt? Browse the archive or send a correction for this guide.