Do you know that feeling of being stuck in a creative process? You are thinking about ideas, but you just can’t make the decisive breakthrough needed? There is a website that will help you get your creative juices flowing. It is called DreamStudio and its’ Betaversion can be accessed via https://beta.dreamstudio.ai. On that website, you can type in any descriptive sentence, like a caption for a photo and click on the “dream-button”. Only some seconds later, you will see four unique pictures that nobody in this world has ever seen before. Even more, the options are endless, even if you process the same text again, you won’t receive the same result. It is created by Artificial Intelligence called Stable Diffusion Generative Adversarial Networks (SD-GANs), with a “text to image” model. Currently the tree most popular models are Stable Diffusion, Midjourney and DALL-E2. In this article we will look more closely at how an image can be created with Stable Diffusion and how the technique behind it works. 

How to generate an image with Stable Diffusion

1. Go to the web application https://beta.dreamstudio.ai
2. Register for an account for free
3. Adjust the options as you like it
   1. Width and Height of the image to be generated
   1. Play around with the Cfg Scale. It adjusts how much the image will be like your descriptive text. The higher the value, the more similar the image looks to your text.
   1. The higher you set the value “Steps” the more denoising steps are spent to generate the picture, resulting in a higher quality. This also influences the time to generate the image. Usually 50 steps are sufficient to generate high-quality images.
   1. In “Number of Images” you can choose how many different images should be generated from the same text. You can choose between 1 and 9. 
4. Type your “Dream” in the text field. Describe the picture as accurate and with as many details, colours and specialties as possible, to get an accurate image
5. Click on “Dream” and wait for your pictures to be generated. 

Costs

After the first registration you will receive credits worth £ 2.  An image generation costs between 1 and 28 credits, depending on the size and how many steps you want to use to generate your image.

So you can create about 200 free images with your first credits. Once your credits are consumed, you can buy additional credits in £ 10 increments. 

How does the technology work?

Stable Diffusion is open source and was released to the public in August 2022. Stable Diffusion is a collaboration between Stabilty AI, RunwayML, EleutherAI, LAION and a research group at LMU Munich. 
It is a type of deep learning that has been trained with a large amount of image and text information, which should produce results as close to reality as possible. More specifically, it is two neural networks, a generator and a discriminator, working together to generate new data that is similar to a given training dataset. The technology of Stable Diffusion adds noise to an image. Then, the noising process is reversed and step bystep improves the quality of the image until there’s no noise anymore. In the case of images, using stable diffusion process helps the generator network to improve the fine-details in the images and make the generated images closer to the real-world images. That is the reason why it’s called “stable diffusion image generation model”.

Stable Diffusion is the latest trend and already widely used for social media filters, art projects and presentations. Let’s see how the technology and use of it will evolve in the coming months but for sure it could be useful for creating realistic images for games and movies, generating images for machine learning training set, creating images for art and design, generating images for architectural visualization or creating images for product visualization. The images can even be used commercially, as stated in their terms.

Author: Giulia Zumstein

The Image was created with dreamstudio. The descriptive text to the picture is: “rising water level in the arctic flooding a little island, dark colors”

Sources: https://towardsdatascience.com/ai-generated-art-how-to-get-started-with-generating-your-own-images-821bc130fef4 accessed on 11.01.2023

https://www.hitc.com/en-gb/2022/12/20/what-is-stable-diffusion-ai-and-how-does-the-image-generator-work/accessed on 11.01.2023