Why AI still can’t recreate brand logos or products—and how that’s changing

For director Craig Kenny, the most difficult part of creating a mock Chanel ad from generative AI was not the 24-hour deadline he’d given himself, nor was it having to shuffle between six different tech platforms to animate the creative from text to image to video. Rather, the heaviest lift was simply crafting the small bottle of perfume at the center of the ad. “Creating the Chanel bottle, in itself, took three-plus hours to design,” Kenny told Ad Age. Kenny’s experience highlights a point of friction that has rankled brands for nearly two years, and one that impedes the potential of generative AI tools for creating ads. While image and video platforms demonstrate an impressive ability to visualize backgrounds or human faces, they continue to have trouble depicting brand products and logos. “Even if you have the most amazing prompting skills, you can never generate what you need,” said Pietro Schirano, co-founder of AI platform EverArt. EverArt is one of a handful of upstart companies at the intersection of marketing and technology hoping to change this. Using various techniques, from fine-tuning an open-source model to building a new one from scratch, these platforms are starting to provide brands a media-generating foundation that is personalized to their style and desired assets. At the same time, tech giants are also trying to help marketers develop personalized content. Meta recently began testing the ability for brands to upload logos to AI-generated images, while Amazon is enabling clients to animate product stills into short video clips. And Adobe has helped agencies rebrand with new characters and styles. Also: How AI is transforming the agency pitch process The problem with using these larger models for specific content generation, such as using them to emulate a brand’s logos and specialized fonts, is that much of the data they’re trained on will be irrelevant, Schirano said. The larger models from the major tech platforms also tend to be closed-source, meaning their underlying code—and training history—is inaccessible. This makes a process like fine-tuning far more constrained, Schirano said. Check out the latest Ad Age events and award programs Fine-tuning, conducted through a process called low-rank adaptation (LoRA), is EverArt’s bread and butter. Schirano analogized it to performing brain surgery on an AI model, in which its training diet is slightly altered to include a small subset of a brand’s data. The resulting brand assets outputs, which EverArt has provided for small fashion retailer Anatomie and e-commerce platform The Naked Market, reflects the details of those assets. “It’s like telling the AI, ‘Anything you know about hats, now all of that knowledge is going to be replaced by knowledge of this one hat,’” Schirano said. In this case the hat is a brand product or design. Using a handful of high-quality brand images or a style guide, EverArt fine-tunes a text-to-image model called Flux. Compared to popular image generators such as Midjourney, Flux allows for easier training because it’s open-source, Schirano said. Recent news: 5 AI takeaways from Advertising Week 2024 AI tends to handle visual complexity quite well, capable of illustrating environments and scenes with almost Bruegel-like descriptiveness. But as soon as text is introduced into an image, such as a logo on a soda can, the image models get confused, Schirano said. Indeed, Kenny, the director who created the mock Chanel ad, told Ad Age he had particular trouble illustrating the letter “C” (for “Chanel”) on the perfume bottle. Fine-tuning is meant to streamline this process and to create in a way that reflects a brand’s bespoke style guidelines. Schirano pointed to the fact that cans of Bubly and Coca-Cola may be similar but require different fine-tuning processes for the model to faithfully represent each brand’s respective look. EverArt often posts its brand product and logo recreations on social media, such as for New Balance shoes or an Oura ring. While the company hasn’t formally partnered with these brands, it has attracted new clients through this approach. “It’s like a cold reach-out,” Schirano said. hi @ouraring our video has gotten pretty good lately pic.twitter.com/xEEb2eRF9a— EverArt (@everartai) October 16, 2024 Models from scratch While LoRA-based fine-tuning is a step up from using generic models online, it may still not get the job done for brands needing perfectly faithful digital assets, according to Darius Lam, founder of AI platform Nex, which develops models from scratch to solve the product and logo conundrum. Lam sees two primary reasons why a “media-foundation model,” which is a foundation that corresponds to a brand’s specific media demands, better suits marketers. The level of control offered by such a model could yield assets more representative of what the brand had in mind, and these assets would be “IP-consistent,” that is, consistently reflective of the brand’s intellectual property and likeness. “Every brand can have access to their own model built on their own IP,” Lam said. More: This tool could help brands monitor for deepfakes Nex, which declined to name brand partners, starts work with clients in much the same way as EverArt, requesting a pool of proprietary data that can be used to accurately portray their products or designs. However, instead of emphasizing ease of use by requesting only a small handful of assets, Nex requests more than 100 assets per product, including images, videos and 3D representations. For any given model, Nex trains it on a catalog of products, such that a brand will end up providing several thousand files representing its identity. Nex trains its proprietary AI model, Ikon, on this information, and develops a system meant specifically for the brand. This system can be continuously trained on new data to improve its generations and reflect additional products. Brands may face a difficult choice choosing a humble model like Ikon over a massively-capable model such as OpenAI’s GPT-4, but Lam sees the technology’s compact size as a strength to counter another AI pitfall affecting marketers: copyright issues. Since Nex trains Ikon from scratch, it is aware of all the data that is hoovered up—an especially important assurance as AI platforms, such as AI search engine Perplexity, face mounting allegations of infringement. Some larger providers including Adobe have also tried to separate themselves from the pack by emphasizing copyright protections, but the size of their models wield less control over copyright than Ikon, Lam said. Brands have started experimenting with media-foundation models. Last month, Lionsgate partnered with AI video startup Runway to develop a model based entirely on the studio’s portfolio of film and TV content, with the ultimate purpose of generating cinematic video. AI is good at handling certain forms of intellectual property—say, Disney’s animated characters—but fumbles more realistic material such as brand products, said Noah Brier, founder of AI marketing organization BrXnd.AI. Startups are using AI video generators to mimic this animated style, as in the case of video AI platform Showrunner, which has created Disney- and “South Park”-esque franchises without using proprietary data from those entities. Cartoonish characters aren’t as intricate, and that makes them easier for AI to portray compared to products and branding set in the real world. In other words, Mickey Mouse is simpler for AI to conjure than a Chanel No. 5 bottle. According to Brier, brands that want to use AI at this point in the technology’s development would do well to emulate the aesthetics of Disney’s IP in their marketing. “The product [depiction] has to be the product, but IP allows more room for error,” Brier said. NextGen Marketing Summit Join us for a crash course on Gen Z, DTC marketing and the creator economy Register today

Why AI still can’t recreate brand logos or products—and how that’s changing

Featured Stories