Fake Volvo ad made with AI stirs marketers—the creator offers 5 tips for using text-to-video tools

A mock video ad for Volvo generated buzz this week for appearing authentic, despite being entirely generated by AI and having no connection to the Volvo brand. Instead, the 46-second spot was developed by Laszlo Gaal, a colorist who used text-to-video platform Runway’s newest model, Gen-3, to portray a Volvo-esque vehicle weaving through a city and spawning lush vegetation in its wake. Gaal used the automaker’s Volvo For Life philanthropic fund as inspiration for the ad. Volvo did not respond to a request for comment. The mock ad, which required less than 24 hours to complete, indicates how quickly text-to-video AI platforms are improving. But while marketers are eager to incorporate this technology into their production processes, they have yet to figure out best practices for generating consumer-ready content. Also read: How text-to-video AI tools can help brands Its apparent believability is largely what made Gaal’s fake ad stand out to creatives and members of the ad industry. “This is one of the most impressive commercial film” made with AI, wrote Matthieu Lorrain, a creative lead at Google’s AI research lab, DeepMind, who posted the video on LinkedIn, where it had garnered nearly 1,000 reactions at the time of writing. “Really impressive … now we have Volvo adverts that aren’t too far off from passable,” commented Justin Westcott, global lead for Edelman’s technology sector. While the mock ad received plenty of praise, others said it lacked storytelling and felt dull. Speaking with Ad Age, Gaal provided his five key tips for marketers hoping to develop convincing ads with text-to-video AI. From the planning phase to prompting and more, here’s how a proven case study breaks it down. See all of the Global Automotive Marketing Awards winners here Know the model’s limitations Before prompting anything, the first step you should take is to identify the model’s limitations, Gaal said. This will prevent you from wasting time and resources by attempting to generate content that the model is incapable of producing. For example, some models are not trained on real-time information, which means a newly birthed concept or item could still be unknown to the model. Other information that may be off the table is more advanced knowledge, such as unorthodox camera angles, which Gaal found Gen-3 incapable of replicating, he said. Technical limitations are important, too. Gen-3 can produce videos no longer than 10 seconds, so if a user wants to create a one-minute spot, multiple generations will need to be stitched together. The final version of Gaal’s ad included footage from 45 different outputs. Understand model-dependent nuances in communication Different models require different styles of prompting, which Gaal likens to the way that people in real life alter their way of speaking depending on their audience. Communicating with a relative differs from communicating with a friend, which differs from communicating with a boss, etc. Some models require a less detailed prompt, such as one sentence, to generate accurate outputs; others, including Gen-3, benefit from roughly five sentences of detail. For his prompts, Gaal first tried using those he had previously inputted for text-to-image platform Midjourney. But in Gen-3, those prompts did not output what he wanted, so he had to adjust the wording and emphasis. “Whenever a new tool comes out, you have to re-learn how to talk to that tool,” said Gaal. Pay attention to hallucinations Once you’ve got the foundation of your prompts down, and your outputs are starting to come together, the key to improving your video is in the hallucinations, Gaal said. Hallucinations refer to the tendency of generative AI models to produce incorrect or unwanted material. Paying attention to where the output deviates from an intended goal, and then finding its reference in the prompt is how you can focus your adjustments, as opposed to re-writing the whole query. But hallucinations need not always be unwanted. They can surface a detail that may inspire you to take the creative to a new place, said Gaal, which he practiced when creating the Volvo ad. In order to take advantage of these opportunities, flexibility is key. “When you have the freedom to follow these different threads, then the end result will be much more interesting than if you have just a storyboard and want to generate 20 very, very strict shots,” Gaal said. Kill your darlings In post-production, Gaal recommends a painful but effective approach: kill your darlings. Text-to-video models can produce so much detailed material so quickly that, before you know it, you could have scores of outputs—each one of which you love for different reasons. But letting the vast majority of these darlings go is a necessary step in the tightening process. For example, the 45 outputs that Gaal used for his final Volvo ad were distilled from a collection of 500. Gaal suggests that marketers try to remember that consumers will never see all the good material that is cut—they will only see the end result. Knowing this can help you to renounce your attachment to outputs no matter how special they seem. Don’t rely on AI for depicting brand products A prominent critique posted in comments to Lorrain’s LinkedIn post is that the actual vehicle in the Volvo ad is inconsistent. Some viewers were unable to determine which model the spot was supposed to be highlighting, and noticed that the angles and interior of the car were unrealistic. Generative AI has become somewhat notorious for committing these failures of fidelity when it comes to detailed depictions. But this is a non-negotiable for most brands, whose ads typically need to spotlight items or other products with pristine clarity. The only current solution for text-to-video platforms, according to Gaal, is to stop using them for these kinds of depictions altogether. Other technologies, such as 3D modeling, can design a far more accurate representation of brand products, which can then be incorporated into the AI-generated video, Gaal said. The process would be similar to how auto brands sometimes shoot spots for vehicles that may not be available yet, in which a dummy car is required for the physical shoot, but is later replaced with media of the new model in post-production. “A hybrid solution between AI-generated films and traditional visual effects—that would make the clients happy at the end,” Gaal said.

Fake Volvo ad made with AI stirs marketers—the creator offers 5 tips for using text-to-video tools

Featured Stories