Oli Yeates · 31 October 2026

AI image generation has grown up. Has your view of it?

In 2021 AI could not draw hands. Every image in this post was generated by AI, with nothing photographed, styled or retouched. Nine releases, one imaginary sportswear shoot, and what it now means for brands.

AI image generation has grown up. Has your view of it?

In January 2021, OpenAI unveiled DALL-E, and the idea of typing a sentence and getting a picture back went mainstream. The early results were fascinating, blurry and frequently unsettling. Less than 6 years on, every image in this post was generated by AI, with nothing photographed, styled or retouched by a person.

If you tried AI image generation in those early days and walked away, here's how far the models have come, and what that now means for brands.

AI-generated sportswear campaign photograph of a basketball player frozen mid-jump against a deep green backdrop, wearing a saffron vest and charcoal shorts, releasing the ball from his fingertips
Generated in Gemini. Frozen mid-air, the ball leaving his fingertips, and 10 fingers exactly where they should be.

Remember the hands?

Cast your mind back to 2021 and 2022, when AI image generation first went mainstream. The results impressed for about 3 seconds, until you counted the fingers. Six-fingered hands, melted faces, gibberish on shop signs and products that looked vaguely like something you once bought.

A recent book on AI captures this moment neatly. Early image models struggled with fingers, and some people concluded too quickly that illustrators had nothing to worry about. Others asked the better question: these tools couldn't do this at all 5 years earlier, so what happens when they improve?

Its authors, Eliezer Yudkowsky and Nate Soares, put it neatly. Those mangled fingers, they argue, were

"...as bad as 'generative art' would ever look."

In other words, the clumsiest output you ever saw from these tools was the floor, not the ceiling. Today they draw hands without a second thought, as every image in this post shows.

Nearly 6 years in 9 releases

The jump from mangled fingers to the photoshoot below didn't happen in one go. It came through a steady run of releases, each fixing something the last got wrong.

Timeline of AI image model releases from DALL-E in January 2021 to GPT Image 2 in April 2026, with each model's mark

The pattern is clear. Hands were largely fixed by 2023 and text by 2025. The past 12 months have been about consistency, resolution and speed: the things that make AI imagery usable in real campaigns.

Why this matters for marketers

Plenty of marketing teams tried AI imagery early, got a hand with too many knuckles or a product that wasn't quite their product, and filed the whole thing under "not ready". That was a sensible call at the time. It isn't a sensible position now.

If your last serious look was 2 or 3 years ago, you're judging today's tools by a version that no longer exists. Current models can produce:

  • photorealistic lifestyle scenes with natural-looking people
  • clean, legible text within images
  • the same product shown accurately across different rooms, angles and seasons
  • high-resolution output, up to 8K, suitable for print and large-format digital
  • short video clips generated from a finished still

The photoshoot that never happened

To put that to the test, we briefed an imaginary sportswear campaign and generated it in Gemini. No athletes, no studio, no photographer, no lighting crew. Each image took minutes rather than a production day.

Look closely at the things AI used to get wrong: hands, skin, sweat and light.

AI-generated sportswear campaign photograph of a female sprinter in a charcoal and saffron kit crouched in starting blocks against a terracotta wall, chalk dust rising from her hands
Chalk dust, sweat on her shoulders and a hand planted on the track. It reads like a commissioned campaign shot.
AI-generated close-up photograph of a female runner after a race, breathing hard, sweat beading on her face, charcoal and saffron vest against a dark terracotta background
Freckles, pores, flyaway hairs and individual beads of sweat. This is the level of detail that used to need a photographer and a retoucher.
AI-generated photograph of the same woman as the runner, now with loose dark hair and an emerald silk dress, laughing at a candlelit dinner table with friends
The same "woman", hours later and a world away. New hair, new clothes, new light, same face. Keeping a person consistent from shot to shot was impossible until 2025; now it’s what makes a whole AI campaign possible.

What's still hard

Making 1 impressive image is now easy. Anyone with a ChatGPT or Gemini account can do it in seconds.

The hard part is making 50 images that are all on brand, all show the right product and all stand up to scrutiny. That's where most brands still get stuck: a kitchen range rendered slightly wrong, a logo in the wrong lockup, a family that looks fine as a thumbnail and odd at full size.

Consistency is what turns AI imagery from a novelty into something you can actually run a campaign on.

We saw it ourselves while preparing this post. Prompts asking for a front row of ordinary, unglamorous faces or a street full of legible signs fell apart. Only simpler briefs, with fewer people and less text, came back usable.

See how far it's come

Judge AI by what it can do today, not by what it got wrong the first time you tried it. If you walked away from AI imagery a while ago, we'd love to show you what's possible now. Get in touch with the team to talk about AI imagery and video for your next campaign.

Sources