Рубрики NewsSoftware

«The most advanced» Stability AI model scared the network with images of mutant humans — but fingers are in place

Опубликовал
Катерина Даньшина

Recently, Stability AI released its new Stable Diffusion 3 Medium image generation model, which it advertised as «the most advanced» — but limitations in the neural network training data led to a problem in reproducing human anatomy.

Stable Diffusion 3 Medium generates creepy images of three-armed, three-legged or headless people. According to social media posts, problems most often arise if the query includes a background on grass or water (but it is unclear what this is about).

The comments suggest that the output of images with «mutant humans» is related to the limited training data of the new Stability AI model, as the company probably censored most of the images with people.

The British startup advertised Stable Diffusion 3 Medium as «the most advanced open source text-to-picture model». It contains 2 billion parameters, is supposed to provide photorealistic results, and eliminate «typical errors on hands and faces» — probably, the problem with the wrong number of fingers was fixed, but as we can see, the overall body anatomy has deteriorated extremely.

Currently, users can test Stable Diffusion 3 Medium via Stability API

Founded in 2020, Stability AI was quickly recognized as one of the new leaders in generative artificial intelligence alongside competitors such as Midjourney and Dall-E by OpenAI. In 2022, investors valued the company at $1 billion. At the same time, already in 2023, Stability AI faced with a number of lawsuits from artists, CEO of Emad Mostak resigned in Marchand the company thought about selling it due to the cash crisis.

Disqus Comments Loading...