If a person becomes less recognizable after several AI photo edits, stop using the latest result as the default input. Find the last version whose identity you approved, and make the next required change from that version. An image can become sharper or more attractive while moving further from the person you meant to preserve. Repeatedly correcting the newest face can also carry its earlier mistakes into later decisions.
Find where the branch went wrong
Lay out the original, the last approved portrait, and the latest result at a comparable face size. Check features that matter for this subject: facial proportions, a characteristic asymmetry, hairline, and any feature the person specifically asked you to retain. Separate these from intentional changes such as clothing or background. Different lighting alone is not evidence that identity changed.
Write the first unacceptable change next to its version. For example, a hypothetical sequence might have an accepted coat change in version B, a narrower jaw in version C, and a successful background replacement in version D. Asking D to restore the jaw is one possible experiment. Returning to B for the background change is another, with a cleaner identity starting point. Neither route guarantees success; they expose different risks.
Keep a small branch record
| Version | Input authority | Allowed change | Identity decision | Next use |
|---|---|---|---|---|
| A | Original portrait | None | Baseline | Preserve |
| B | A | Coat only | Approved after inspection | Eligible master |
| C | B | Background only | Jaw changed: reject | Do not feed forward |
| D | B | Revised background request | Pending | Review before reuse |
This is a planning example, not a completed generation test. Keep the actual files alongside the record. A filename such as final-final does not tell a collaborator which version supplied the face or which change was accepted.
Split preservation from invention
For the next attempt, name the trusted image and the limited request together. A useful brief is: “Use the approved portrait as the identity and clothing authority. Change only the setting to a quiet courtyard. Preserve the face proportions and the accepted coat. Do not inherit facial alterations from the rejected example.” Only attach the rejected image if the tool needs it for a clearly limited purpose; otherwise it can create another competing authority.
An existing successful result can legitimately become a new reference. Runway documents this kind of iterative workflow. The decision here is whether that particular result has passed your own identity check, not whether all iteration is inherently harmful. Keeping untouched source data and separate versions also follows the practical principle behind non-destructive editing.
Decide when to stop
Approve the requested change and the preserved identity separately. If a localized repair is available and suitable, consider it before regenerating the whole portrait. If neither branch preserves the person, reduce the requested transformation or return to a clearer authorized source. Do not disguise a changed face by adding skin detail. For an Imagild workflow, keep the chosen input and final download so the next edit begins from an explicit, reviewed version.