10 Photos That Show What AI Image Generators Struggle With Most [CNET]

One of the things AI image generators struggle with the most is photorealistic images. When I prompted Leonardo with “1980s girl band performing on stage,” I wanted something that would resemble a photograph, and the images were way too oversaturated and fake-looking. And I wasn’t quite envisioning the droopy, vampiric facial characteristics, either.

This image was made by what we ranked as our top AI image generator, OpenAI’s Dall-E 3. It’s compelling, but if you look closely, you’ll see a cat, a two-headed dog and other problems. Even the best AI programs aren’t always right.

This is a great example of an amazing idea that didn’t end up exactly right. I was impressed that Leonardo knew enough about TMNT and da Vinci’s famous painting to mimic their styles and colors, but when turning humans into turtles, some characteristics got wonky. Points for trying!

This is one of the funniest images Canva’s Magic Media AI tool created. Accurate facial expressions was one of Canva’s biggest stumbling blocks, but the pure horror and exaggerated fear on these people’s faces made me laugh. Human expressions continually challenge AI image generators.

I really have no idea what’s happening in this one, and I don’t think Midjourney knew either. I used Midjourney’s built-in editing tools to expand one of the original photos, and instead of adding more soccer players, it added… whatever that is on the right. Midjourney gets points for having editing functions, but sometimes they make things worse.

In defense of Canva here, none of the AIs I tested were able to render specific company logos well. This is a great example of how even well-known brand icons can’t be recreated with AI, no matter how much you adjust your prompt.

While reviewing Leonardo, I found that one of the things it was best at was creating beautiful scenery. Unfortunately, the library room of my dreams is ruined by the rolling ladder to nowhere. This is a classic case of disappearing overlapping objects.

This Canva image looks fine from afar, but when you look closer, the details unravel. Check out the lettering on the cookbook page and the blurriness of the cutting board in the middle ground. Images like these highlight how frustrating it can be to create AI images. Even when you think you’ve got a winner, a small flaw could render it unusable.

Dall-E 3 has a conversational tone and can help you improve your prompts. In this case, it really amplified the prompt and the overly dramatic end result missed the intention of the original prompt. That’s why good AI image prompt engineering is so important.

A lot of problems can come up when generators try to create real-looking people. Changing the style to cartoon or anime-looking can occasionally help, but as you can see from this woman with many, many fingers on her hand, it doesn’t always work.