Artificial intelligence, or “AI” for short, is on everyone’s lips. Image creation is particularly popular with many people. Corresponding programs are currently springing up like mushrooms, and the internet and social networks are seemingly flooded with AI-generated images and clips. It is becoming increasingly difficult to maintain an overview of this enormously growing field. Prof. Dr. René Peinl, Head of the Institute for Information Systems at Hof University of Applied Sciences (iisys), is an expert in the field of artificial intelligence. He reveals which programs and applications you really need to know about!
Midjourney – the market leader
“The program works in a somewhat unusual way. It uses a chat server (Discord) to send requests to the model. If you don’t pay anything, your requests are placed at the back of the queue and processed with low priority, i.e. very slowly. This means that you have to wait several minutes. As a paying customer, you have a quota that is processed within a few seconds. In terms of quality, the program is currently the reference, albeit with a slight gap to its competitors.”
Examples of images created with “Midjourney”:
DALL-E 3 from OpenAI – precise with content requirements
“The program is most easily accessible via the Bing Image Creator. DALL-E 3 is also of a high standard, especially when it comes to the implementation of content specifications. However, if you want photo-realistic results, you will often be disappointed. The results look more like well-made, realistic drawings. On the other hand, even complex prompts (short instructions, e.g. in the form of general questions or precise instructions for execution) are usually implemented in great detail. Other styles such as Impressionism or the style of Salvador Dali are also mastered well. The advantage: you can generate a few images a day for free via the Bing Image Creator if you have a free Microsoft account.”
Examples of images created with “DALL-E 3”:
Stable Diffusion – for friends of OpenSource models
“I am a self-confessed advocate of digital sovereignty and therefore a supporter of open source models. Stable Diffusion from Stability AI is an equal competitor in this area. Stability AI’s basic models themselves are good, but not quite at the same level as Midjourney or DALL-E 3. Stability AI itself recently started offering a cloud service with its own models for a fee.
Recently (28.11.23) there is a version SDXL turbo, which is similarly good as SDXL, but five to ten times faster. With all the alternatives mentioned (including Midjourney and DALL-E), images are created in several steps by gradually reshaping random pixel patterns to produce the desired result. With SDXL you need 20-40 steps for good results in high resolution. With SDXL turbo it is only 5-10. Smaller images can even be created in 2-3 iterations in decent quality.”
Examples of images created with SDXL:
Fooocus – the uncomplicated program for beginners
“There are also dozens of free third-party programs that further refine the basic model. Many of them deal with very specific use cases, e.g. photorealistic images of people or animals, fantasy paintings, etc. These can be installed on your own PC. You can install these on your own PC without needing a degree in computer science. A good example is “Fooocus”. The program is explicitly aimed at beginners. The settings are deliberately kept somewhat hidden. But you get good results without detailed knowledge of prompt engineering, i.e. formulating image generation requests. The AI implements them in the best possible quality 1:1. If you want to go deeper, you can still change a lot in the settings and achieve even better results.”
automatic 1111 – for advanced users
“This program is also easy to install, but does not hide the dozens of setting options, so that it can initially seem a little overwhelming for non-experts. However, if you want to try out different models, this is the best way to get started. The relevant settings are often documented by the creator of the model, so you only have to transfer the values without understanding exactly what they mean in detail.
However, Fooocus and automatic 1111 only provide the user interface and the “trappings”. For both, you still need the image generation model “under the hood”. However, one is supplied to get you started straight away, others are available on Huggingface, for example.”
By the way…
Hof University of Applied Sciences is also hosting an image generation model at the Institute for Information Systems (iisys), which will be available to all university staff and students from the summer semester of 2024. It is currently still being tested.
Disclaimer: Hof University of Applied Sciences has no connection whatsoever with the companies mentioned.
(Status: 16.01.2024)
