An overseas brand commissions a household-product film from a production partner in China. An AI mood board gives the imagined lead a polished face, perfect light and an effortless smile. When real auditions arrive, the discussion becomes “none of them looks like the reference.” The team may be evaluating distance from an image rather than ability to perform the role.
For live-action advertising, ONCE proposes a role-before-reference review. Define the performance task, record an initial assessment of real candidates, and then discuss the generated image for its declared visual purpose. This is a production method to try, not a demonstrated way to improve campaign results.
A research signal with a clear limit
The October 15, 2026 preprint Generative AI Produces and Transmits Beauty Bias audits 27 image generators. Its human-rating study found generated faces more attractive than the photographic comparisons under the evaluators' standards. Such ratings are judgments, not an objective scale of human beauty.
A separate hiring simulation found that exposure to deliberately selected attractive AI faces lowered subsequent evaluations of identical real candidates. The authors caution that this involved one job and one session, and does not establish an effect for ordinary output from every generator. The paper was checked on October 27.
It did not test advertising casting, actor performance or sales. The useful production question is narrower: what comparison environment are we creating before evaluating real people? A casting review should preserve room to describe what a candidate actually does before asking how closely they resemble a synthetic ideal.
Write a role card without a portrait
Start with the situation, objective and product interaction. What is the character trying to finish? What changes their response? Which language or practical skill is genuinely needed? Add appearance and styling requirements where the creative gives a reason for them, rather than reverse-engineering requirements from an attractive generated image.
Imagine a hypothetical vacuum-cleaner commercial in which a visitor arrives earlier than expected. A useful audition task might involve hearing the doorbell, checking the remaining floor area and deciding what to do next. It does not require a claim about a particular product's speed or performance.
“Busy homeowner” is a starting description. The decision made after the doorbell is an observable performance. The character may be glamorous or understated; the team should know why that choice belongs in this story. A model's default polish is not itself a casting requirement.
Keep the review in four stages
| Stage | Review material | Record the decision |
|---|---|---|
| Define | Role card and actual scene | Required action, language and relationship |
| Observe | Authorized candidate material and a comparable task | The moment that works and the adjustment needed |
| Reference | Generated visuals with a stated purpose | Lighting, wardrobe or composition to explore |
| Reconcile | Role, audition notes and reference purpose | Whether to change casting, styling or the reference |
This order is a proposed precaution, not proof that bias disappears. If reviewers have already seen the AI image, acknowledge it. Do not pretend that hiding it now resets everyone's judgment. Instead, ask for rejection reasons tied to the role and a visible passage.
A note such as “the performer continues the rehearsed action after the doorbell” gives the director something to address. “Less premium than the reference” does not. Where candidate photographs use different lighting, avoid attributing every visual difference to the person. Agree on suitable comparable conditions if another review is needed.
Make the image's job explicit
Attach a short use note: “Reference for side lighting and wardrobe color; not a match target for facial structure, skin finish or expression.” If appearance is part of the creative, express the requirement in words with a reason and let the casting team assess a realistic search. Do not promise an exact human counterpart to an invented face.
Then decide which object needs changing. A candidate may perform the required hesitation while the board depicts a permanent smile; the board may need revision. A wardrobe mismatch belongs in styling. An unclear product action may require direction or a different shot. These are distinct decisions, rather than reasons to keep making the actor look more polished.
Do not alter a real candidate's photograph toward the synthetic reference and circulate it without clear agreement. Identify real material, generated concepts and post-production previews as different things, and secure appropriate permission for any personal material used in the project.
Commission exploration and casting separately
An AI exploration package can deliver visual directions and styling possibilities. Live-action casting involves candidate material, auditions and performance judgments, followed by availability, willingness and usage arrangements. Approval of the generated portrait does not lock a real performer.
Specify the exploration rounds, audition passage, candidate comparisons and adjustment scope in the quotation. If the project changes to a synthetic lead, revisit how the performance will be made and delivered. Treating that change as a simple face replacement can conceal substantial production decisions; no universal saving is assumed here.
For an ongoing content program, retain role tasks, reference purposes and specific performance observations where permission allows. Those records are more useful than a permanent attractiveness ranking of real people. Begin the next supplier brief with a role card without a face, then decide what the reference image should help explain.
For a different casting question—how two performers respond to each other—explore ONCE's broader Video Production articles. This article addresses the comparison standard introduced before an audition, not chemistry or a tested audience response.
