Leonardo tests GPT-Image 2.0 on ad briefs

For the past couple of years, the entire tech industry evaluated image generation models using the exact same playbook: Did it get the fingers right? Can it spell words on a sign? Does a crowded street turn into a blurry mess?

Those are valid technical benchmarks, but they miss the point of how real work actually gets done. Most people who end up using these tools have zero prompt engineering skills. They don't want to type arcane camera settings; they want to drop in a plain idea and get back something that looks like an experienced creative director touched it.

Leonardo.Ai's Head of Creative, Dwayne Koh, put OpenAI's latest GPT-Image 2.0 to the test on ad briefs to see what happens when you strip away the clever prompting tricks. Koh, whose portfolio spans campaigns for adidas Originals, Levi's, Jordan, Vans, and Converse, wanted to know if raw briefs could yield production-grade work.

The viral sticker nobody asked for

Instead of crafting a hyper-detailed prompt, Koh fed the model a basic ad scenario: a fashion-forward celebrity launching a glow lip balm, tailored to a 14 to 30 audience, with visuals that resonate with Gen Z and young millennials.

What came back surprised him. The composition nailed the low camera angle to give the subject presence, but the real shocker was a subtle detail: GPT-Image 2.0 spontaneously slapped a "viral on TikTok" sticker onto the ad layout.

Nobody asked for it in the brief. Yet in the world of beauty marketing, that tiny badge is precisely the cultural shorthand a human art director would reach for to build hype. The model wasn't merely converting words to pixels; it was interpreting the audience and making an unprompted creative decision.

In a second test for a fictional streetwear brand called Threads, Koh asked for a cool group shot of friends hanging out. The model automatically introduced dynamic smoke bomb effects—a visual trick Koh noted his team frequently used at adidas Originals to elevate campaign mood boards. Competing models given the identical prompt landed nowhere near production-ready standards.

Why it matters

We think this test flags an important shift in synthetic media: AI image models are graduating from dumb rendering engines into taste-reflecting collaborators.

For solo creators, small marketing teams, and startup founders, this changes the game. You no longer need to learn complex prompt syntax to extract high-end aesthetics. If you can clearly articulate your audience and brand tone in everyday language, the model can deliver solid visual concepts that would otherwise require days of back-and-forth with an agency.

For the broader industry, it redraws the battle lines between frontier labs. Once technical hurdles like rendering text and anatomy become table stakes, the real moat is cultural comprehension—whether a model understands irony, street culture, typography hierarchy, and subcultural cues.

That said, having good taste on the first pass doesn't mean AI is ready to replace human art directors. A spontaneous TikTok sticker is a brilliant spark for an early-stage mood board, but refining that output into finalized, multi-channel ad deliverables still requires rigorous editing and tight brand control that machines can't yet manage alone.