AI has finally figured out how to draw hands – removing the one easy way to identify deepfake images
One of the biggest giveaways for spotting a deepfake image may have just been fixed after a leading artificial intelligence image generator figured out how to create realistic hands.
The latest version of Midjourney, which generates images from simple text prompts, marks a significant improvement from previous iterations of generative AI technology, which typically depicts human hands as oddly formed or with the wrong number of fingers.
Midjourney V5 features “much higher image quality, more diverse outputs, wider stylistic range, support for seamless textures, wider aspect ratios, better image prompting [and] wider dynamic range], according to its creators.
Improvements included more realistic skin textures, better reflections and shadows, and more detailed facial expressions.
But users were quick to notice that the biggest upgrade appeared to relate to the AI’s ability to generate human-like hands.
Human hands are considered by many artists as among the trickiest things to draw, and early AI image makers from Midjourney, Stable Diffusion and OpenAI’s Dall-E showed that generative artificial intelligence also struggled.
— Weird Ai Generations (@weirddalle) January 22, 2023
It meant that otherwise convincing images of poltical figures like Donald Trump and Emmanuel Macron could be recognised as fake just by looking at the hands.
“Just a headsup – Midjourney’s AI can now do hands correctly,” said video game artist Del Walker.
“Be extra critical of any political imagery, especially photography, you see online that is trying to incite a reaction.”
AI researchers have previously attempted to explain why generative artificial intelligence trained on billions of images from the internet struggle with certain aspects of the human form.
“It’s generally understood that within AI datasets, human images display hands less visibly than they do faces,” a spokesperson for Stability AI told BuzzFeed News earlier this year.
“Hands also tend to be much smaller in the source images, as they are relatively rarely visible in large form.”