Case Study 01

Character Consistency Across Real-World Scenes

From a single reference image, I built the same character across varied real-world environments while preserving identity, realism and natural variation.

One reference → Six generated scenes → Identity consistency

Role: Concept, art direction, generation, visual QA
Discipline: Generative image
Focus: Identity consistency, photorealism, prompt iteration
Output: Six finished scenes from one reference

Generated range

Six environments, one person — every frame built from the same reference.

Generated scene: the character laughing in a café

Generated scene: the character inside a convenience store

Generated scene: the character beside a train window

Generated scene: the character walking on a city street

Generated scene: the character at a table in a restaurant

Generated scene: the character taking a mirror selfie in a lift

Consistency logic

What was held, and what was allowed to move.

Reference portrait used as the identity anchor

Held constant

Facial structureProportion, eye spacing, nose and jaw line across every scene.
HairLength, colour, part and the way it falls under different light.
Skin characterTexture, tone and apparent age, rather than a retouched surface.
Body proportionBuild and shoulder line, so scale reads the same at any distance.

Allowed to vary

LightCafé warmth, store fluorescents and daylight through glass.
ExpressionLaughing, neutral, distracted — read as the same person reacting.
FramingDistance, lens character and crop change with each environment.
StylingWardrobe shifts with the setting without becoming a costume change.

The reference carries no environment information. Everything else in the series — light, wardrobe, location, framing, expression — is built after that point, which is what makes the drift measurable.

Process

Lock the reference and write the identity description that travels with every prompt.
Build each environment separately — light source, surfaces, time of day, camera distance.
Generate in small batches per scene so identity drift is visible early instead of at the end.
Compare each candidate back to the reference, not to the previous generation.
Refine the prompt where the face slipped; regenerate rather than repair after the fact.
Select the frame where identity, light logic and expression all hold at once.

Checked on every candidate

Face drift between scenes — proportion, eye spacing, jaw
Age drift under softer or harder light
Hairline and hair volume continuity
Skin reading as skin, not as a smoothed surface
Hand and finger construction
Light direction agreeing with the environment
Reflections and glass behaving physically
Background text and objects holding together

Outcome

Six scenes that read as photographs of one person on six different days — the basis for any campaign, series or character-led content that has to stay recognisable across formats.