2 min Applications

Stable Diffusion 3 is a step back for AI images of humans

Stable Diffusion 3 is a step back for AI images of humans

Recently, Stable Diffusion 3 Medium was released. This caused much hilarity due to the model’s difficulty in correctly representing a human body. As a result, images surfaced online of people with three legs, crazy fingers, and arms that merge into legs.

Mocking posts about Stable Diffusion 3 Medium appear online; one user even asks, “Is this release supposed to be a joke?” Compared to other creators of AI graphics, such as DALL-E 3, this third release of Stable Diffusion is a miss.

The glaring images shared online show that the model has trouble rendering the human body even slightly correctly. In other types of images, the model will most likely function properly.

The release of Stable Diffusion 3 Medium is actually a step backwards for the company. Earlier models, such as the SD XL Turbo, which was released in November, did not exhibit these problems. Stable Diffusion 3 launched in preview in February. This model would be released in several models, each with a unique size between 800 million and eight billion parameters. The recently released medium model contains 2 billion parameters.

Targeting adult content too harshly

There is speculation online about the cause of the problems. Most think it is due to a filter. Since the first release of tools for AI images, companies have had to take more action against 18+ content.

Research has already been conducted on the effects of such filters on previous models of Stability AI. On this, researchers at AssemblyAI wrote: “Recall that 2.0 is trained on a subset of the LAION dataset that has been filtered for inappropriate content using an NSFW filter, which in turn results in a relatively reduced ability to depict humans.” With stricter filters, the consequences are more clearly visible.

Also read: Stability AI exec steps down, citing concerns over generative AI ‘exploitation’