One Character, Four Locations, Eight Revisions: An Identity-Drift Test

Character consistency is the quiet killer of AI-assisted campaigns. A face that looks right in the first frame can age, soften, or subtly rearrange its features by the fourth location and the sixth revision note. I subjected Gemini Nano Banana 2.1 to a structured identity-drift test: one character, four distinct locations, eight sequential revision requests that mirror the notes a client or creative director actually sends. The goal was not to find a perfect generation. The goal was to map exactly where likeness, wardrobe logic, and expression continuity begin to erode under commercial pressure.

Constructing the Identity-Drift Protocol

The Character Lock

The character was defined with the same rigor used for product markers: approximate age range, face shape, distinctive features (a slight asymmetry in the smile, specific eyebrow density, hair length and parting), skin tone, and a simple wardrobe baseline—light linen shirt, minimal jewelry. These markers were written into every instruction as non-negotiable. Any frame that altered them failed the likeness check even if the new image felt more polished.

Four Locations That Force Environmental Change

Location one: bright interior café window light. Location two: shaded outdoor market stall. Location three: soft evening interior with practical lamps. Location four: overcast coastal boardwalk. Each location required new lighting logic and supporting environment while the character remained the same physical person.

Eight Revision Requests That Mimic Real Notes

1. Keep character identical, change only the background season cues.

2. Soften the expression slightly while protecting all facial markers.

3. Change wardrobe from linen shirt to light sweater, keep face and hair identical.

4. Adjust body angle fifteen degrees, preserve likeness.

5. Add a simple prop (ceramic cup) without altering character scale or features.

6. Make the light cooler while locking every facial and wardrobe marker.

7. Tighten crop to three-quarter portrait, keep identity stable.

8. Final request: “She looks great—can we make her feel a touch more approachable without changing who she is?”

Drift Observations Across the Sequence

Early Locations Held Likeness Cleanly

The café and market frames preserved face shape, smile asymmetry, and hair parting with only minor lighting-driven shifts in skin rendering. Wardrobe baseline stayed consistent. At this stage a client could have approved the first two frames as the same person in different environments.

Expression and Wardrobe Revisions Introduced Softening

The “soften expression” request produced a plausible new expression yet slightly rounded the jawline and reduced the distinctive smile asymmetry. The sweater change preserved face better than expected, but hair volume increased subtly—enough to notice in a side-by-side strip. These were not catastrophic failures; they were the first measurable identity costs of revision.

Later Revisions Compounded Small Errors

By the cooler-light and tighter-crop requests, the cumulative effect of earlier micro-shifts became visible. The face still read as related, yet a strict likeness board would have flagged it. The final “more approachable” request—broad aesthetic language—produced the largest single drift: a smoother, more generic face that lost the original asymmetry. Broad language remains the fastest way to erode a locked identity.

Practical Continuity Controls That Improved Results

Protection-First Instruction Language

Instructions that began with “Keep the exact face shape, smile asymmetry, eyebrow density, hair parting, and skin tone identical. Change only [variable]” consistently outperformed instructions that led with the desired change. The model still made interpretive decisions, but the magnitude of drift decreased.

Side-by-Side Strip Evaluation

Evaluating each new frame in isolation hides drift. Placing the original locked portrait beside every revision at identical scale makes even small changes obvious. This simple QA step caught problems that thumbnail review missed.

Knowing When to Stop Iterating Inside the Model

After the sixth revision the cost of further generation exceeded the cost of finishing remaining notes in a design or compositing tool. The lighting and environmental ideas were already strong. Restoring exact likeness markers became a precision task better suited to tools that can reference the original face directly.

Campaign-Level Implications for Character Work

Usable Range Versus Locked Range

Under current behavior the model can deliver two to three strong, identity-consistent frames with careful prompting and QA. Pushing to four locations and eight revisions without human likeness restoration produces a set that feels related rather than continuous. For social campaigns that tolerate mild variation this may be acceptable. For any use case that requires the same recognizable person across a full sequence, plan for design-tool or compositing intervention by the midpoint of the revision list.

Wardrobe and Prop Changes Are Safer Than Expression or “Feel” Requests

Concrete wardrobe or prop swaps introduced less facial drift than soft expression changes or vague tonal requests. When the brief demands an emotional shift, the safer path is often to generate the new expression as a separate controlled test and then composite or carefully blend rather than asking the model to reinterpret the entire face.

Documentation That Travels With the Asset

Every character sequence should carry a short identity board: the locked hero portrait plus call-outs for the five or six markers that must not move. Future collaborators—or future you—can then evaluate new frames against the same standard instead of relying on memory or aesthetic impression.

One character, four locations, and eight revisions remain a stringent but realistic commercial test. Gemini Nano Banana 2.1 can hold likeness longer than many earlier tools under narrow protection language. It still accumulates identity drift when revision count and environmental change both rise. The professional response is not to abandon the tool; it is to treat likeness as a protected system that requires the same continuous QA a traditional casting and continuity team would apply on a multi-day shoot.

For AI tool features, rules and availability, refer to their official sites.