An advertising image can become increasingly correct. The subject is clearer, the palette is easier to like, the setting looks more relevant to the audience, and pretest scores become more stable. Yet when ten brands iterate toward similar high-scoring examples, something else happens: their pictures start to resemble one another, while the one product truth worth remembering becomes less visible.

This is not evidence that generative AI is inherently uncreative, or that optimization inevitably damages brands. The problem is more precise: a model faithfully amplifies the objective it receives, and that objective usually measures what is easy to accept more completely than what could belong only to this brand.

A Model Does Not Optimize a Good Ad. It Optimizes the Metric in the Loop

A 2026 issue of the Journal of Marketing published Picture Perfect: Engaging Customers with Visual Generative AI. The researchers connected consumer feedback directly to image generation. They collected advertisements and audience ratings, then fine-tuned an open-source image model on AIDA mindset metrics, brand personality, and segment-level feedback. In their automotive, sunscreen, and related experimental settings, consumer-feedback-trained visuals could match or outperform conventional ads on the metrics being optimized.

The important change is not simply that AI can draw. It is the feedback loop:

  1. Define an objective such as attention, interest, desire, or activation.
  2. Ask real people to evaluate existing visuals.
  3. Train the system on the relationship between visual examples and those ratings.
  4. Generate more variants that move toward the chosen objective.

This is more useful than prompting alone because it introduces who finds the image effective. It also creates an editorial problem. If the highest-scoring examples share safe compositions, familiar settings, and category conventions, the model learns not only effectiveness but also an average.

The Visual Average Is Effective, Not Neutral

The authors describe the result as a visual average built from elements associated with successful advertising. That does not mean low quality. Averaging can increase processing fluency: people understand the image faster, and familiarity, trust, and liking become easier to evoke.

In one supplementary test, 998 participants rated ads, and generated visuals produced higher processing fluency than conventional ones, 5.17 versus 4.80. Fluency also correlated with AIDA performance. A generative system can therefore be very good at removing comprehension friction.

Brand distinction, however, sometimes depends on friction. An unusual proportion, an unexpected product situation, or a precisely timed joke may feel less immediately comfortable while creating stronger ownership.

The paper's Smart Fortwo study provides a useful boundary. The two-seat car has a distinctive promise built around compact size, city parking, and agility. Its conventional advertising used surprising situations to dramatize that difference. Images trained on high-performing mainstream automotive visuals produced slightly lower average AIDA ratings than actual Smart ads, 3.62 versus 3.81, with p=.064. This is not a decisive statistical failure, but the direction and mechanism matter: the visual language that works for mainstream cars can make a distinctive car look ordinary.

Humor produced a clearer limitation. The researchers trained on both high-AIDA and highly rated humorous dog-care ads, yet could not generate above-average humor. In a later memory test, the real humorous ad achieved higher brand recognition than the generated version. Humor depends on contextual incongruity. It establishes an expectation and then breaks it at the right moment. An average is good at establishing the expectation; it may be poor at deciding which deviation could belong to the brand.

Route Production by the Cost of Becoming Average

The useful division is not AI versus live action. It is whether a task can safely move toward the average.

The Scale Layer: Turn Familiarity Into Production Efficiency

The scale layer suits work that can be evaluated clearly, requires many versions, and does not need a new creative concept each time. Examples include:

  • adapting one product language across aspect ratios, backgrounds, and audience contexts;
  • extending an established brand system into social assets, site banners, and regional combinations;
  • refining color, information density, and settings through real audience feedback;
  • producing catalog variations, functional contexts, and lightweight motion studies with stable display logic.

Here, the average is useful. It reduces unproductive variation and preserves time for decisions that need human judgment. Google's own Demand Gen image-generation guidance keeps advertiser expertise, asset selection, and final approval in the loop, while recommending ongoing testing and asset reporting. Generation expands the option space; it does not decide what the brand should become.

The Deviation Layer: Write the Frame That Must Not Look Average

The deviation layer carries brand memory, unusual product positioning, humor, surprise, and a new visual grammar. It should not begin with a request for ten high-scoring images. It should begin with an anti-average proposition:

  • If every category ad puts the product in a wide landscape, should the space be barely large enough to contain this one?
  • If every character succeeds immediately, should this story preserve a hesitation, a wrong judgment, or a second attempt?
  • If competitors show the whole product first, should this brand introduce only a distinctive action or sound?
  • If pretests favor warmth, brightness, and symmetry, does one justified asymmetry need to carry the memory?

These choices require the creative lead, director, art department, cinematographer, and brand team to explain why the deviation is necessary. Generative tools can still explore storyboards, situational variants, lighting, and material hypotheses. Their job is to help the deviation become visible, not to vote it out with a larger pile of average alternatives.

The Reality Layer: Let Products, People, and Physics Face the World

Still-image findings do not automatically generalize to every film, performance, or product demonstration. Video adds continuity, causality, duration, sound, and acting. A high-scoring key frame is not the same as a believable film.

The reality layer protects what cannot be replaced by looking plausible: real material behavior, functional actions, contact between people and products, sound-image causality, manufacturing or use processes, and any outcome the audience may interpret as fact.

The NIST Generative AI Profile places generative systems inside a full lifecycle of design, development, use, and evaluation. For advertising teams, the practical inference is that generated output needs versioning, provenance, and human review before it becomes ordinary production material.

Our article on AI and sustainable advertising made a related distinction: imagination can be generated; organizational responsibility cannot. Product visuals can also be extended, but product facts still need real objects, records, and appropriate testing.

A Better Brief Defines What Must Be Familiar and What Must Not Be Averaged

A brief made only of words such as young, technological, international, and premium invites an error-free category average. Divide the objective into two columns.

The first is what must be familiar. What should the audience understand immediately? Which use situation, brand color, information hierarchy, character state, and channel rule must remain stable? These elements can be scaled through data, templates, and feedback.

The second is what must not become average. Which product truth, character choice, compositional relationship, action rhythm, or sonic cue must differ from category convention? If the logo disappears, what still allows the image to be attributed to this brand? Lock this element before scaling generation, or optimization will repeatedly pull it back into familiar territory.

Our work on brand content spacing explains why core assets need to return across content cycles. Generation is well suited to varying their contexts, but the recognition core should not be redefined by the highest-clicking version every time. Otherwise, the team gains more assets while losing the memory those assets could have accumulated.

Do Not Select One Winning Score. Read Three Curves

Review candidate executions on three separate curves.

The first is comprehension. Can people recognize the product, situation, and primary information quickly? Fluency is valuable here.

The second is attribution. With the logo removed, can people assign the image to the brand rather than only to the category? This is the test of visual ownership.

The third is deviation. What breaks expectation, and does it create useful memory, meaningless oddity, or factual risk? Interviews, real audience testing, and creative judgment should explain that response together. A single metric is not enough.

The goal is not to maximize all three curves in every asset. The scale layer can prioritize comprehension. The deviation layer can prioritize attribution and memory. The reality layer can protect factual integrity. A content system can then contain many dependable versions and a smaller number of irreplaceable brand images.

The Average Is Not the Enemy. It Just Cannot Make the Final Brand Decision

The commercial value of generative AI comes from turning feedback into repeatable visual production. Its creative risk comes from the same mechanism: repeatable objectives pull difference toward a shared center.

Brands do not need to choose between generating everything and rejecting generation. A more mature system lets models scale familiarity, lets people create justified deviation, and lets live action and verification protect reality. The element worth defending in a creative review is not every unusual idea. It is the difference that may not have the highest immediate score, but still explains the product, belongs to the brand, and survives real viewing.

If you are building a generative visual system, a brand content matrix, or a hybrid AIGC and live-action workflow, bring ONCE the target metrics, product distinction, existing high-performing assets, and the brand cues that must never be averaged away. We can design an AIGC commercial and hybrid production route that separates the scale, deviation, and reality layers before deciding what to automate.

Sources and Scope

The cited research focuses mainly on static visuals, selected product categories, defined participant samples, and explicit evaluation tasks. This article translates those mechanisms into creative and production decisions; it does not treat them as conversion guarantees for every brand, market, or video format.

← AI Disclosure in Sustainable Ads: When Warmth Backfires AIGC Content Systems: Why Failed Generations Matter →