A fake Dior ad posted to LinkedIn last month became the latest AI-generated dupe to catch marketers’ attention. But this was no deep fake, rather a transparent attempt by a production company to push the bounds of what is possible when using AI to create advertising.

The 50-second spot, titled “The Ascendance” and developed by Serbia-based Davinci Production, tried what many other AI experiments seek to avoid, such as representing brand products and logos and depicting real people (in this case, Rihanna and actress Emilia Clarke). This was part of an intentional effort to go beyond demonstrating AI’s ability to create striking visuals, which has mostly been the focus of creatives’ attention, said David Blagojevic, co-founder and film director at Davinci Production. 

The result was a 29-day exercise that illustrates the creative possibilities that may be awaiting marketers on the other side of experimenting with generative technology.

Davinci’s ad builds on previous explorations by creatives using real brands as templates for their AI-generated content. Over the summer, a colorist went viral for creating an impressive Volvo ad, which combined the typical style of an auto spot with thematic imagery inspired by Volvo’s real-life philanthropic fund. And video AI startups have generated longer forms of media meant to resemble legacy franchises, including those owned by Disney. 

As with its precursors, Davinci created the ad without the actual brand’s permission. Blagojevic, who acknowledged his company’s lack of association with Dior in his initial LinkedIn post, declined to comment on whether the luxury brand has since contacted his company.

Dior did not respond to a request for comment.

Also read: How AI is transforming the agency pitch process

As a departure from other recent specs, Davinci’s was less concerned with mimicking a typical spot than fleshing out areas that have yet to be established for AI-generated advertising. One such area is ad format. Similar to how influencer marketing on social media has developed formats in terms of length, style and other characteristics, AI-generated content may very well require its own standards to find a lasting place within marketers’ strategies, Blagojevic said. 

“Everybody knows that we can do a striking shot [with AI], so something new will have to happen for it to keep being interesting,” he said.

The format used by Davinci was similar to a TV commercial but with noticeable stylistic differences. Firstly, the ad does not follow a coherent storyline. Instead, Blagojevic wanted to use AI’s abilities in illustration to capture the look and feel of the product being advertised, Dior Prestige La Crème. The luxury skin cream’s bottle has a golden cap, which informed much of the recurring colors and items depicted in the spot.

Another important aspect of the ad is its departure from realism. Generative AI has been criticized for its tendency to create unrealistic assets, often marked by a noticeable glossy sheen. In Blagojevic’s view, this is not a problem without a precedent. Unrealism can be found in all types of conventional ads, from depictions of clothing to detergents that wash them. What may be needed, then, is to leverage AI in a way that still effectively promotes the product and its corresponding consumer experience, Blagojevic said.

“It will be very obvious that commercials aren’t real, but does it actually make [the product] not real?” Blagojevic said.

Experimenting with real depictions

Another boundary Davinci tried to push was depicting things usually kept off-limits, such as real people. This has been a key controversy in the generative AI space, and one that brands, without proper permissions, have mostly avoided leveraging in their advertising.

Given that Davinci had set out to show what is possible with AI technology, the company decided to charge ahead.

“I said, ‘Okay, we need that,’ because everything changes once you see somebody who you know on the screen AI-generated,” Blagojevic said.

Depicting real people is not only a risky endeavor in terms of legality and because doing so is difficult, and viewers can quickly identify any inaccuracies. This is partly why brands including Unilever’s Dove have pledged not to use AI to portray human features in their advertising.

More on AI: 9 AI tools agencies are embracing right now

Davinci’s approach did not consist of hammering away at an image generator like Midjourney until it produced close enough resemblances to Rihanna and Clarke. Rather, it fine-tuned open-source models using low-rank adaptation (LoRA). This method of fine-tuning, or training a pre-existing model on a specific data set, consisted of Davinci uploading images of the celebrities so that the underlying AI system could learn the likenesses and create them. Davinci applied the LoRA training via the Fal and Replicate platforms, which work on top of the open-source Flux model created by Black Forest Labs. 

The process is not as easy as drag-and-drop, however, and even if the AI can faithfully recreate a face that does not mean it will do justice to the rest of the figure.

“More often than not, the little details are the most challenging ones,” Blagojevic said. “You set up the whole story, and then you find you have six toes on a model’s foot.”

Davinci also used LoRA to create the Dior product at the center of the ad, the same tack being taken by several AI martech startups. This presented difficulties because each angle from which the product was shown needed its own lifelike recreation, Blagojevic said. This required training the model on roughly 30 different images of the skin cream bottle to allow for continuity once the bottle was animated into video, which Davinci enabled through the image-to-video platform Runway.

Davinci’s depictions drew some ire from viewers. In a comment on the original LinkedIn post, Christy Nunns, a video producer and founder of Wavelength Studio, suggested an apparent hypocrisy between Blagojevic acknowledging the need for talent to have rights over their likenesses and his creating and sharing the ad anyway.

Not only was it “creepy” that Davinci recreated real people without their consent, but it points to a larger problem between AI and creativity, Nunns told Ad Age—namely, that rampant AI use is being conducted without thinking about potential consequences. One possible outcome of the Dior ad, Nunns said, is that could spread online and erode trust in the brand as well as its advertising. This flow of misinformation could happen regardless of whatever acknowledgments were initially attached to the post.

As a creative is judged by the worst of their work, so AI will be judged by the worst thing it’s associated with, Nunns said.

Art direction leads the way

While Davinci wanted to embrace certain aspects of AI’s stylistic uniqueness, it did not want to create an ad that felt too obviously computer-generated. This is the downside to the democratization of creativity that AI has enabled: everyone is using the same tools, and thus creating very similar content.

Homogeneity presents an especially problematic issue for marketers because their ads need to stand out to successfully promote real products.

“We cannot make a film that lasts for one minute and have all the shots look like they were filmed at the same locations with the same actors and the same product,” Blagojevic said.

One way to push past this obstacle is by allowing human art direction, not AI, to take the front seat, Blagojevic said. Just as a real director would guide the vision of a non-AI ad, so too should they guide an AI-generated one, the thinking goes.

Blagojevic recommends that brands hire directors who have direct experience with creating ads for the category being advertised. Blagojevic himself is a veteran of cosmetics advertising, so he felt comfortable guiding the AI for the Dior ad. This point is a caution against marketers who think that they can now create effective ads because they have generative AI at their fingertips. To the contrary, brands need to continue hiring creative teams specialized in production, many of whom are now experimenting with AI, Blagojevic said.

One telling example is the technical know-how that Davinci still needed to employ in order to finalize its mock Dior ad. The company generated over 13,000 shots during its 29-day project—far more than a non-AI ad would yield. But the final product contains less than 50 shots. All of that trimming into a coherent spot is where experienced art direction comes in handy.

And so, too, with the touch-ups that will inevitably be needed due to AI’s imperfections. Davinci had to use graphics software Inpaint to edit numerous shots, Blagojevic said, but in a way that did not make the non-AI content contrast with the AI-generated content. This act became tricky with respect to backgrounds, which Davinci had a hard time animating with an image-to-video platform. Instead, it opted to use a green screen and non-AI content to construct this part of the ad, before laying it back in behind the AI-generated elements.

“You have to be sneaky about this stuff, because you always need something more that is not possible to do with the current technology,” Blagojevic said.