Meanwhile, the market for text-to-image AI has exploded. More and more groups are trying to get their share. I actually wanted to write an article about tips for making good text prompts, but I couldn't see the wood for the trees anymore. So here's an attempt to sort things out.
What have we:
- The big AI developers
- Providers of Open Source AI
- Platforms to develop AI on
- Providers for the general public
- AI prompt generators
- organization for safe use
- websites about AI art (like this one)
This is probably not complete, but it should be enough to get an overview.
The big developers
The big AI developers occasionally make the papers, and then the courtrooms. These AI developers have small and large research departments that are constantly developing better AI. I limit myself here to the major AI art generators:
- OpenAI (Dall-e, Dall-e2), once founded by Elon Musk c.s. to stimulate the development of AI, now in close relationship with Microsoft. ChatGTP is also theirs and is based on the same text comprehension algorithm.
- Stability AI (Stable Diffusion version 1.x and 2.x), founded by Emad Mostaque, has Amazon Web Services as a partner. They are focused on making AI accessible with Open Source, not just for generating images.
- Google (Imagen) is known by everyone on the internet. Imagen is not yet widely touted, but Google is a major player in the AI field. Especially since they took over Deepmind (belongs to Google since 2016), although Deepmind focuses on 'more useful' applications.
- Midjourney (Midjourney) calls itself "an independent research lab exploring new ways of thinking and expanding the imagination of the human species." It is self-financed and is led by David Holz. The tool is available through Discord.
- NVIDIA is best known as a maker of graphics cards with accompanying software, but let that be exactly the best hardware for AI. (For the insiders: both revolve around matrix calculations.) NVIDIA's tools that apply AI seem focused on rendering 3D images, but their research is broader.
Of course, research is also being done at universities worldwide into the latest applications of AI, where they can collaborate with one of the major players above.
In principle, OpenAI, NVIDIA and Google do not use Open Sourcing of their tools, while Stability AI does. As a result, anyone can now (in principle) create, train and use these advanced neural networks themselves. The research of all parties is partly public. For example, Károly Zsolnai-Fehér's "Two Minute Papers" shows in short YouTube videos the latest research papers in the field of AI (in English), with the best results and occasionally also the bloopers. This also includes research results from Google and NVIDIA. In addition to releasing the source code and research, there is another way to give every citizen access to the AI: Most of these companies have a portal on their website where you can try them out for free and on a small scale.
Thanks to the open sourcing of the latest algorithms, everyone can in principle build and train a neural network on their own PC, but this takes a lot of computer capacity and therefore time. Now you can also visit platforms such as Weights&Biases (wandb.ai) and Hugging Face (huggingface.co). They offer an environment for young and old to build and test their own Machine Learning algorithm. This is where the development of new AI is facilitated. If you want to train an algorithm on a limited set of art, so that you only generate this art movement, you can get started here.
For the general public
Most of the picture generators mentioned above can be used commercially. You often pay per picture by first buying an amount of credits, and handing in one or more credits per picture.
DALL-E2 can be used via commercial providers, such as NightCafe. OpenAI lets you try it for free on its own website, but for big work and commercial use it offers access via an API.
Midjourney offers Midjourney through Discord. The first 25 creations are free, after that you have to pay. Other organizations can install the Midjourney bot in their own Discord environment and draw their own audience there.
BlueWillow calls itself an aggregator of multiple AI models, including models such as Stable Diffusion. BlueWillow also works via Discord. The number of followers of this is increasing day by day, so I will not list them here. These startups often start for free, until they have developed themselves enough to be able to ask for money.
Stability AI offers the trained Stable Diffusion algorithm on open source platforms such as GitHub and Hugging Face. It also offers APIs so that the AI can be used within Photoshop, for example. Those who do not want to play with code themselves can go to Stability AI's DreamStudio. There you pay for the use (and sponsor further research). You can also use Stable Diffusion via NightCafe. Simple assignments are free, for complicated assignments you pay with credits.
Lexica.art also offers the Stable Diffusion algorithm, with a 'Free plan' you get 100 free images per month. You make them with 4 at a time, so you quickly reach that limit.
The platforms on which you can generate images give users plenty of opportunities to look at each others art. On some platforms you have to take a more expensive subscription to keep your work hidden, with the others it is a free choice. A large proportion of users consciously show their prompts so that others can learn and they can improve together.
In addition, websites have emerged that want to generate a prompt for you, or compare for you, or sell to you. The latter seems nonsensical, but is probably an option for someone in an incredible hurry (and with plenty of money). I want to dedicate a next blog to all of these websites.
Organization for safe use
From 2016, the development of AI started to gain momentum, but the first (already predicted) problems also became visible. That is why the Partnership on AI was founded, for responsible AI. It's not yet really visible to the general public, but fully active. I had to discover them myself via Wikipedia: the (in full) Partnership on Artificial Intelligence to Benefit People was officially presented on September 28, 2016, with founders Amazon, Facebook, Google, DeepMind, Microsoft, and IBM. Apple joined in January 2017, and by 2019, more than 100 partners from academia, civil society, industry, and nonprofits joined the coalition. In October 2018, Baidu became the first Chinese company to join the coalition. In November 2020, the Partnership on AI announced their AI incident Database (AIID): a database of real-life problems, so developers can try to avoid those situations in the future.
So now you know where to go with reports of abuses with AI, complaints about biases in the datasets they were trained on, or complaints about misuse of the algorithms.
AI art websites
In addition, many - really many - people have started working with the AI to conjure up images from their computers themselves. They can be used for leisure or for work. For inspiration or for therapy. Both from the AI angle and from the art angle people have blogged and vlogged about it, like me on Harmanna-AI.nl. In addition, there are also wikipedia pages, subReddits, and other forums where people meet like-minded people.
If you want to stand out with your website, choose the .ai extension for your domain. This extension comes from the island of Aguilla and is sold for a lot of money. The administration of this British overseas territory in the Caribbean Sea will not regret it.
I think I have my overview back, so on to the next blog.