Consistent Characters and Models Across a Campaign

How to keep a generated person recognisably the same across dozens of images, why drift happens, and the disclosure question that comes with it.

A campaign that features a person needs that person to be the same person in every frame. In conventional photography this is automatic. In generative production it is one of the harder technical problems, because each image is produced independently and the model has no memory of what it generated an hour ago. Faces drift, ages shift, features rearrange, and a set that should read as one campaign reads as several.

Drift is not random and understanding where it comes from makes it manageable. It concentrates in the features that carry identity: the precise geometry of the eyes and the space between them, the nose, the jawline, and the specific asymmetries that make a face individual. Broader attributes such as hair colour, apparent age and build are easier to hold. This means a character can be consistently described and still not be consistently recognisable, which is why description alone is insufficient.

The methods that actually work all involve conditioning rather than description. Establishing a definitive reference set of the character from several angles, then using those images as structural input for every subsequent frame, holds identity far more reliably than any written specification. Where the workflow supports it, training a lightweight adaptation on a consistent set of that character produces the strongest results, at the cost of the setup time required to build it.

Building the reference set properly is the step that determines everything downstream. It should cover the character from front, three quarter and profile, in neutral lighting, with a neutral expression, plus a small number of expression and lighting variants. This set becomes the character's canonical definition, and it should be approved by the client before any campaign imagery is produced, because changing the character later invalidates everything made from it.

Wardrobe, styling and continuity details need their own written specification, since they drift as readily as faces and are more noticeable to viewers. If a character wears a specific jacket in a campaign, the colour, cut, collar and details should be recorded and referenced rather than re-described. Audiences notice a changing garment more reliably than they notice a slightly changed cheekbone, because clothing is what they are consciously looking at.

Consistency of light and lens must accompany consistency of identity, or the character will read as different even when the face is correct. A campaign should fix the key light direction, the quality of that light, the apparent focal length and the depth of field, and hold them across the set. Grading the finished set as a unit rather than image by image corrects the residual drift that survives everything else.

The disclosure question is unavoidable in this category and deserves a settled position rather than improvisation. Kirk and Givi (2025) found that perceptions of AI authorship shape consumer responses to marketing communications and can produce negative responses in some conditions, and Farooq and de Vreese (2026) documented how awareness of AI generation affects authenticity judgements. Wang and Zhang (2025), studying virtual influencers, found that how a virtual figure's identity is framed affects purchase intention, which suggests that presentation and framing matter rather than only the fact of synthesis.

There is a firm legal boundary alongside the perceptual one. A generated character must not resemble an identifiable real person, and any workflow that starts from photographs of a real individual without their consent is a serious problem rather than a shortcut. The U.S. Copyright Office (2024) addressed the digital replicas question directly in its report on that subject, and the practical position for a commercial studio is straightforward: build characters that are not anyone, document that they are not derived from a specific person, and obtain proper releases whenever a real person is involved at any stage.

The honest assessment of where this technique fits is that it works well for stylised, illustrative and clearly constructed campaign imagery, and works poorly where the audience is being asked to accept the person as a genuine customer, employee or endorser. A generated model in an aspirational lifestyle frame is a familiar advertising convention rendered by different means. A generated person presented as a real satisfied customer is a fabricated testimonial, and no amount of technical quality makes that acceptable.

For clients the practical guidance is to decide the disclosure position before commissioning, approve the character reference set formally, and expect the studio to maintain and reuse it rather than regenerate the character each time. A studio that cannot show you the canonical reference set for a campaign character is generating a new person for every image and hoping nobody looks closely, which for a campaign running across a year is a reasonable thing to expect someone to do.

References

Kirk, C. P., & Givi, J. (2025). The AI-authorship effect: Understanding authenticity, moral disgust, and consumer responses to AI-generated marketing communications. Journal of Business Research, 186, Article 114984. https://doi.org/10.1016/j.jbusres.2024.114984

Farooq, A., & de Vreese, C. (2026). Deciphering authenticity in the age of AI: How AI-generated disinformation images and AI detection tools influence judgements of authenticity. AI & Society, 41(1), 493–504. https://doi.org/10.1007/s00146-025-02416-5

Wang, X., & Zhang, Y. (2025). Product-independent or product-dependent: The impact of virtual influencers’ primed identity on purchase intention. Journal of Retailing and Consumer Services, 84, Article 104088. https://doi.org/10.1016/j.jretconser.2024.104088

U.S. Copyright Office. (2024). Copyright and artificial intelligence, Part 1: Digital replicas. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-1-Digital-Replicas-Report.pdf