One of the first AI models that could generate images painted this market. Four years later, I asked today’s technology to paint the same scene again.
The comparison shows how quickly image generation changed.
Why faces and hands kept falling apart
The old image came from the early mass-market phase of image generation, around 2022. Back then, models from Hugging Face still ran on gaming PCs before Midjourney made access easier.
Many of those systems used diffusion: they gradually removed noise and assembled statistically plausible patterns from training data. There was no explicit 3D model of the scene and no anatomy check.
Faces exposed the weakness immediately. Eyes, noses, and mouths have to align precisely, so small errors make a face look wrong. Hands were even harder because of their many poses, fingers, overlaps, and occlusion. In training images, they also often appeared small or cropped.
The models could imitate texture without reliably knowing where fingers began and ended. More resolution usually only made the mistakes sharper: faces melted, hands negotiated their own anatomy, and text turned into nonsense.
Click either original to open it in full resolution.
Four years later
The newer version keeps the market, shelves, people, and trolley, but materials, expressions, and composition are far more convincing.
Generated text remains difficult. The most obvious anatomical failures, however, have largely disappeared. In just four years, “obviously generated” became an image that needs a second look.
For children, generated images will sit beside photographs, films, and illustrations as something normal. With every image, the important question becomes where it came from and what we are meant to believe.
Perhaps one day we will choose alternative film endings generated in real time with the same characters and voices.
That is why I am cautious about predictions. The distance from distorted faces to believable people was just four years. Wahnsinn.
What is the oldest AI-generated image you still have, and what would you create if you could ask today’s models to remake it?