O Stable Diffusion now it’s 2.0. The new version of artificial intelligence (AI) that generates images from texts brings new features, such as increased resolution. But not everyone is happy. The tool’s ability to generate nude images and imitate the style of certain artists has been reduced.
Complaints started pouring in just a few hours after the announcement of Stable Diffusion 2.0🇧🇷 At the reddit, there were even those who claimed that the tool was “nerfed”. This is a way of saying that the update was weaker than the previous version, despite bringing new features.
What’s bad (according to complaints)?
Perhaps the biggest novelty of the new version of the tool is OpenCLIP, a new encoder to “convert” text into an image. Developed by Laion with support from Stability AI (the company behind Stable Diffussion), the feature generates images with much higher quality compared to the previous version of the tool.
But this is where the complaints start. The text templates linked to OpenCLIP contain filters that remove nudity or pornographic images.
This limitation, in itself, is considered negative by many people. But the restrictions go further. Some users have noticed that the new version also filters instructions to imitate the styles of certain artists.
One example comes from the work of renowned illustrator Greg Rutkowski, who creates beautiful, painterly images. It is no longer possible to reproduce his style in Stable Diffusion, not with faithful results.
Users who have noticed the restriction believe that this is a way to prevent Stability AI from facing legal problems. Asked about it by VergeEmad Mostaque, founder of the company, did not respond.
However, the businessman explained that, contrary to what users speculate, the new version of the AI did not remove images of artists from its training database. According to him, there were only changes in the way Stable Diffusion encodes and retrieves data.
The mention of Greg Rutkowski it is not mere chance. The artist has already made illustrations for games like Horizon Forbidden West and year🇧🇷 It’s such a remarkable piece of work that its style has become popular with users of tools like Stable Diffusion.
Rutkowski is one of many artists who are uncomfortable with this new scenario. At first, he thought that the tools of artificial intelligence could help you reach new audiences. But he changed his mind when he found a work that bore his name and style but was computer generated.
“That was just a month ago. But what about in a year? I probably won’t be able to find my work out there because [a internet] will be flooded with AI art.” vented Rutkowski In September.
The subject has given rise to much debate. In general, there is concern that AI-generated images devalue the work of a community that, with the exception of established artists, already struggled to gain recognition.
Up until anime and manga artists are already concerned therefore.
What has improved in Stable Diffusion 2.0?
Arguments aside, Stable Diffusion 2.0 has also been garnering praise. That’s because the novelty got much “smarter”. Merit of its new features.
By default, the tool generates images with a resolution of up to 768×768 pixels. But the new version brings an upscaling mechanism that expands this capacity to 2048×2048 pixels or, in some circumstances, even higher resolutions. The direct effect is images with much more detail.
Another advancement is the depth2img feature. Through it, Stable Diffussion 2.0 uses the depth data of an image to generate others with the same parameters, but adding the instructions that the user gave by text.
Note the example below. In it, the original image is that of the gentleman without glasses. The images that appear in the sequence show different people, but with the same profile as that gentleman. The AI already had this capability, but only generated an image as a result. Now there are several.
A new inpainting model that lets you swap parts of an image easily and optimization for Stable Diffussion 2.0 to run on a single GPU are among the other new features.
This last point is relevant because the Stable Diffusion is open source🇧🇷 Thus, the tool can be implemented by any person or organization.
https://tecnoblog.net/noticias/2022/11/25/stable-diffusion-irrita-usuarios-ao-impedir-imagens-com-estilo-de-artistas/