Last year, few generative AI companies released an improved public version of their image generators including Midjourney, Stable Diffusion, and DALL-E, among others. These programs are using deep learning based on billions of pictures made by humans to generate images from natural language descriptions. They enable people to create digital images quickly in a way no human is capable of, making combination of things beyond our imagination. The generated images can represent fantasies of all types, ranging from landscapes never seen before to surreal characters and scenery.
Some tools such as Midjourney and Stable Diffusion require the users to enter a detailed text called a prompt which describes the image that will be created. Once the prompt is submitted, the tool provides several image variants which can be used or refined in multiple iterations. Other tools such as Artbreeder allow users to create new images by remixing existing ones, or change image details of a portrait such as color balance, character gender, age, and apparel. These alterations within a generated image are called inpainting. In contrast, users can experiment with outpainting, which allow expanding an image beyond its original border by adding visual elements in the same style or similar scenery.
As suggested before, artists can use AI generators in their work to create images that will eventually become paintings. While sophisticated in its ability to produce visual elements via algorithm, the AI generator is like a paintbrush. It is to a digital artist what a camera is to a photographer. So, they have the potential in making artists better at what they do instead of making any user a professional artist. Some artists working in illustration and design might use AI generators as a brainstorming tool in the early stages of projects. While other artists might use the generators to quickly create prototypes or drafts of artistic concepts to discuss with their clients before starting their work.
Aside from their practicality, it is difficult to engage with some AI-generated images beyond appreciating their novelty. Digital artists often use editing tools to refine their creative images, questioning the ability of AI-powered tools to be creative. With time the AI tools might become more powerful and accurate, and will eventually generate even better images. However, people will still need to be creative whether by finding the right prompts to describe their idea into the image generator or by using a generated image as a basis for their final artwork.
In October 2018, Christie’s sold an AI-generated painting titled Edmond de Belamy (2018) for $432,500 in New York. The painting belongs to a series of eleven generative portraits called La Famille de Belamy published by Obvious, a Paris-based art collective working with AI to create art. The portrait represents a blurry image of a man in a black coat facing the viewer, reminiscent of neoclassical painting. In fact, it was generated through a long training process that involved around 15,000 portraits painted between the 14th and the 19th century.
In August 2022, an AI image generated with Midjourney, titled Théâtre d'Opéra Spatial, won a contest for digital art organized by the Colorado State Fair. The image printed on a canvas was well received by the jury of the art fair since the artist didn’t break any rule when creating his work. However, many artists rejected the painting because it was an AI-generated artwork. They thought an image produced by AI was not art because the tools lacked the creative spark and skills supplied by humans. More artists are seeing AI generators as a threat because they are explicitly trained on existing work of artists.
Furthermore, most of the AI companies behind the text-to-image tools claim no rights on generated images. They let people use any generated images as long as their content is not illegal. For digital artists, this policy represents stolen labor as their works are obtained without their consent, and might lead to lost jobs or opportunity. In January 2023, several visual artists filed a class action lawsuit against generative AI companies. Similarly, Getty Images pursued legal action against Stability AI for training millions of their images using Stable Diffusion. The company had previously banned AI-generated images, especially produced by DALL-E, from its platform.
A major concern that might be related to such a ban is the proliferation of deepfake – digital images manipulated to replace facial expression and appearance through deep learning. To avoid this risk, the AI generators are configured to reject any prompt involving public figures. They can also analyze uploaded images to detect any offensive material or images containing familiar faces.
However, this precautionary measure has led to another concern around algorithmic bias related to gender, race, and age. Indeed, OpenAI engineers found that removing violent and sexual imagery ended up reducing the frequency of women in the generated images. They deduced that women were more likely to be sexualised in training data, resulting in the introduction of a gender bias. To mitigate that bias, they have configured DALL-E to automatically add gender keywords to prompts that have none.