German enthusiast Fabian Stelzer used online tools based on the popular Stable Diffusion neural network to edit his selfies in different styles. He popularised such algorithmic systems and made them viral, Motherboard.
Stelzer used programs that repeatedly apply user adjustments to the text-to-image generator.
He created selfies in various painting styles, film genres, and ideas such as photographs from “Woodstock” or in the image of a woman.
Stable Diffusion from the startup Stability AI — one of many AI systems that generate images on demand. The developers trained the algorithm on billions of annotated and labelled images.
The neural network can be trained to associate specific words and phrases with certain types of images, aesthetics, places or objects. This has given rise to websites that let users to ‘tune’ the model and add their own embeddings.
The creators of the service drawanyone claim that their tool can process a photo from five initial shots in just an hour. Several people demonstrated their AI selfies in a thread under Stelzer’s post.
According to the enthusiast, he was impressed by the quality of the system, as only a handful of reference images were needed to embed himself in the model.
“The selfies I uploaded may be somewhat similar, so I will retrain the tool on a more diverse set of poses and expressions,” he said.
Stelzer is also involved in other AI projects, including the development of a feature-length multi-plot film, “Salt.” He uses algorithms such as Stable Diffusion, Midjourney and DALL-E 2 to craft frames and a model for generating sound, including Synthesia and Murf. The screenplay is written by the GPT-3 neural network.
As each fragment of the film is released, viewers can vote on one of the continuation options.
“AI systems are at least as significant as the invention of photography or cinema, and if we include language models like GPT-3, then this could be compared with the invention of the printing press. The impact on mass media, culture and the fabric of reality will be fairly profound,” said Stelzer.
The enthusiast noted that Stability AI’s decision to publish the source code will make the algorithm safer for closed corporate projects.
“These systems are also used by bad actors. However, with such projects, fewer users are able to deploy the tool competently,” he added.
In October, the enthusiast used Stable Diffusion to create ever-changing virtual worlds.
In the same month, an AI researcher taught the neural network to turn anyone into Pokémon.
In August, Stability AI provided access to Stable Diffusion to more than 1,000 developers, and later opened it to a wider audience.
Subscribe to ForkLog AI news on Telegram: ForkLog AI — all the news from the world of AI!