AI Is Not a Threat to Real Artists

“AI is a powerful tool, but it’s still a tool. It is not, at least for now, an author of anything. It is essentially an amalgamator,” wrote the artist David Salle, discussing how he trained a proprietary AI model on images of classic painters and his own work to rework a series of old paintings and create his ‘New Pastorals’. “Silkscreen, video, neon – all radical departures at one time, now simply part of the artist’s toolbox,” he also wrote, pragmatically considering how artists experiment, embrace, and reject – as they see fit – the tools that become available to them.

The AI conversation in advertising is more complex and driven by different factors than the one in the realm of AI in art. They are, however, intrinsically connected when we look at the sphere of image production. We’ve already seen the vast improvements in the kind of imagery that generative AI can produce; we’ve seen many of those images integrated into film and print (to mixed reception); and we can be sure that we’ll see it used more widely with time.

The Role of Generative AI in Visual Effects

Jordi Bares, founder and creative director of artist-led VFX studio Rohtau, is keenly aware of how AI is impacting the industry and the kind of images making it to our screens. He stresses that AI’s use in VFX isn’t new; in fact, it’s quite old, and routine to boot. “Tools that optimise algorithms and make things faster – for example, compression algorithms, fluid simulations, lighting visualisation and denoising tools,” Jordi explains.

It’s generative AI, “tools that generate pixels”, that is new, and so is its integration into VFX.

Jordi first started using generative AI around six years ago to aid in style transfers. “Copying the style of an image into another to get closer to the grade or the stylistic approach. I used it for simple little things.” Since then, its capability has evolved, enabling Jordi to carry out more sophisticated tasks like face swaps and full environment generation.

Above: Rohtau's work with digital influencer, Lil Miquela highlighting cutting-edge technology in action.

Human Creativity vs. Machine Generation

Crucially, Jordi and the Rohtau artists never use generative AI to create final images. But models like Midjourney are used during ideation. “We use them to explore ideas as if it were a Google search and then recreate them by hand,” he says. When, after a reference, the team says it’s easier to generate one with AI than contend with the noise and SEO manipulation present in a regular Google image search – it’s become essentially useless, according to Rohtau. “We recently worked on a project where we carried out explorations in Midjourney, and then we executed it in 3D, basically rebuilding everything from scratch.”

“Our general approach is to use it as a reference point. Everyone we know now begins with AI as a way to search better and faster.” Nevertheless, as Jordi has already touched on, that’s only a starting point for an artist, and the road ahead is far more complex than the misconception typically held by clients: input prompt, receive image.

Sometimes it takes more than one model to get close to what’s prompted. “There is no one button to press to see something magical happen. Certainly, when you need control over elements of escalating complexity, we find that we have to mix and match different models with CGI objects. Prompts have to be mixed with example images, with Photoshop sketches, mixed until you’ve crafted that image,” explains Jordi, clarifying how much artist input is needed for AI to ‘generate’ an image.

Rohtau’s head of production and sustainability, Josh King, adds that what clients need to understand is that AI is not “a go faster, go cheaper button”, saying, “It takes craft, it takes time, and in many cases, it's not the best tool to do the job.” According to Josh, generative AI likewise doesn’t give clients the one thing they want above all else: control. “Clients are in a traditional mindset, so they think making tweaks is easy. They ask to change an element or two, but we’re working with generative AI, if we do it again, it will be different - but things are moving fast and consistency is coming… rapidly.”

A point of concern for Jordi is the data used to train AI models, with implications for intellectual property breaches. That’s why when Rohtau employs AI for image generation, the team trains models from scratch. Jordi reveals that he’s currently building “a massive data set” of 260,000 images to train a model. The images are from an actress who was filmed on special conditions with image rights established prior to the shoot – necessary precautions to protect everyone involved. “By doing so, we don't expose our clients to any liabilities because we have everything 100% under control.”

The Craft and Aesthetic Issue

The issue is also applicable to the style of generated images. Anyone who has used generative AI to create images, or seen ones created by it, knows that they have a certain look: airbrushed, saturated colour, and mildly uncanny even when no obvious glitches are present. “What the models create has a very distinct look. And you need to fight against it, actually,” Jordi says. To his expert artist eye, “currently/yesterday lighting tends to be really, really bad and boring. You won’t see beautiful shadows or certain details like a painter would include.”

Jordi sees these generated images as “a race to mediocrity”; a strikingly similar sentiment to the one expressed by Lucky Generals chief creative director Damien Le Castrec, who said that “AI will take everyone to the same place faster.”

What’s more, AI is running out of data. Having scraped all the available images, it’s now training on its own output in a process called a ‘synthetic augmentation loop’ or ‘fully synthetic loop’. And that’s a problem because it can lead to model collapse, degrading the quality of the models’ output. To put it plainly, if we think that an image generated by AI today – trained on original paintings and photography from the past – looks bad, just wait until it starts generating images based on what AI has made. “It will become an echo chamber,” adds Jordi.

“From a craft point of view, the people who are able to create interesting and unique images, real artists, are the ones who will be in demand because AI will never touch them,” Jordi says. Craft, skill, and artistry will be the differentiator in the visual marketplace as screens become flooded with monostylistic images, further desensitising our attention. “We will see a lot more noise, a lot more garbage everywhere, and I believe it will be very counterproductive. Audiences will become exhausted, like they have with Marvel movies,” he says.

What’s needed now, according to Jordi and Josh, is a renewed investment in craft and the artists who can utilise the tools at their disposal to deliver the desired outcome, to develop a certain aesthetic with its human imperfections as well, and to make audiences pay attention. For Josh, “it's about bringing [AI] tools into our current tool set, and then making unique images together. That’s how we can construct images that are much more directable and attention-worthy than gen-AI.”

Meanwhile, Jordi says that he “doesn’t feel threatened at all” by AI’s ‘artistic’ capability. Why? It’s like David Salle wrote, AI isn’t “an author of anything. It is essentially an amalgamator.”