Images created by Stable Diffusion are amazing, but its open source worries

The story of image generators that use Artificial Intelligence (AI) has gained another chapter (with several footnotes): the stable diffusion. It's like DALL-E – which became famous on social media this year for being able to create weird scenes using text commands – only with two important differences: it's already been released to the public and it's open source.

These two points, combined with Stable Diffusion's potential to generate spectacular – and, more importantly, believable – images have generated debates and concerns on the web. Some are excited about the tool's creative possibilities and democratic reach. On the other hand, there are those who are worried about ulterior motives when using it. And some are pissed off because the generator can mimic, almost perfectly, artistic styles (including those who are still alive).

Fact is, Stable Diffusion is riding the wave of creative tools that use deep learning (deep learning) to create visual media, along with other tools such as OpenAI's DALL-E 2; Midjourney; and Google's Image. And because it was just released, your universe still has that lawless earth energy.

What is Stable Diffusion?

It is an AI-powered image generator. With it, anyone with a computer and a relatively good graphics card (at the moment, it only works with Nvidia cards) can create virtually any digital painting they can imagine. Van Gogh's portrait of Beyonce? It's possible. view of a city cyberpunk in the style of the 18th century Japanese artist Hokusai? It can be done.

Yes, the tool can mimic almost any visual style, from famous cartoonists to 19th century daguerreotypists (it really existed, you can Google it to check). And you can also include descriptive phrases and images for reference when configuring the demand in the tool.

Stable Diffusion left the mind of Emad Mostaque, founder and CEO of startup Stability AI, whose mission is to create open source AI projects. It was under this pretext that the new tool came to the public at the end of August. But this is not the first undertaking of the entrepreneur, who is approaching 40 years old.

In 2019, Mostaque, originally from Bangladesh but currently based in England, founded Symmitree. The purpose of this startup, created 14 years after he graduated from Oxford University, was to reduce the cost of technology for people in poverty. A beautiful project, but that did not go forward, discontinued the year after its creation. And then, in late 2020, came Stability AI.

How to use the AI ​​imager?

You can play, for free, with a demo version of Stable Diffusion, which works in browsers even, accessing pages like HugginFace (partner company of startup in the development of the tool), dream studio (commercial page created by Stability AI for the tool) and Google Collaboration. However, all of them are in English, with not very intuitive interfaces. And you have to be patient, because it takes a while for you to get the tool to generate nice images. On the DreamStudio page you can also download and install the tool.

Regarding ways to use the image generator with AI, some defend the idea of ​​working with it as if it were an instrument, similar to a music producer or plastic artist, who chooses different tools when assembling a beat or paint a picture.

Currently, there are more than one AI image generator and each has its strengths and weaknesses. So why not combine them to create something? That's what director Karen Cheng tried to do. She came up with the idea of ​​using AI to generate pieces of clothing and overlay them in a video – and showed the entire process in a thread on Twitter (in this case, she used DALL-E and EbSynth).

Open source and other issues

Okay, so far we've seen the potential of Stable Diffusion (and generators like that) for creative projects. But the world is not one big Disneyland. Much less the internet. So, it is necessary to highlight the thorny side of Stability AI's attitude in releasing such a powerful tool in such an open and almost unrestricted way.

The launch of the tool to the public left many on alert for fears related to cultural and economic impacts. It's just that, unlike Dall-e, the Stable Diffusion database is available for anyone to dig through and use without too many restrictions.

In practice, this means that people can use Stable Diffusion to generate images that OpenAI slashes in Dall-e, for example. Among these types of images are violent scenes, pornography, scenes that may violate corporate and copyright rights, deepfake of celebrities and production of material for disinformation and fake news.

The “official version” of the tool does have filters that block, for example, nudity. However, due to open source, these barriers can be overcome. And considering that Stable Diffusion is there for everyone who wants to use it, controlling (and blocking) use with hidden motivations is basically impossible. Regarding this, the CEO of Stability AI only said:

We believe in individual responsibility. We've included a usage policy and tools to mitigate damage.

Emad Mostaque, founder and CEO of startup Stability AI, in an interview with Art Technica

Another front that has generated a stir is Stable Diffusion's ability to imitate artists' styles. The point is that the company “fed” the tool with the works of the artists without asking for permission or making any kind of agreement. This raised a number of questions related to ethics and copyright.

Next steps for the company

In any case, the success of Stable Diffusion, even though it was launched less than a month ago, has already boosted the valuation of Stability AI in the market. The company even had meetings with investors to raise US$ 100 million (R$ 523 million, at the current price) in resources. And an investment firm would have expressed interest in closing a deal that would raise the startup's value to $500 million, according to Forbes sources.

The company's CEO said, in a recent interview with the engineering machine learning algorithm and YouTuber Yannic Kilcher, who is targeting partnerships with “leading governments and institutions”.

Mostaque also shared, on a social network, that Stability AI invested US$ 600 thousand (R$ 3,1 million) to “train” Stable Diffusion. So far, the investment has paid off.

See also:

DALL-E, the AI ​​that creates images, has public tests released e DALL-E Now Imagine What's Beyond the Borders of Artwork

Sources: Ars Technica, Hackaday, AI stability, TechCrunch, Forbes e DevOps