What is AI generated art?
By Jamie Li
September 10, 2022
UPDATED 12:00 PM EST
Prompt: An astronaut riding a horse in a photorealistic style [Photo Credit: OpenAI]
You may have seen uncanny images that were generated from text floating around the internet. Today, we’ll take a look at text-to-image generated art, DALL-E 2, and the positive and negative implications.
[Photo Credit: OpenAI]
How do these text-to-image AIs function? For a more in-depth look, DALL-E 2, created by Open AI in January 2022, is a great place to start. First, the CLIP model takes images with captions and creates mental representations with vectors. The Prior model takes a CLIP text embedding and generates CLIP image embeddings. Finally, the Decoder diffusion model, also known as unCLIP, takes the CLIP image embedding and generates images. A visual of how DALL-E functions are pictured above.
What are some of the implications of AI-generated art? Firstly, AI-generated artwork greatly reduces the friction that comes with creating art by lowering costs and time. For example, an aspiring musician can make their album cover, game designers can create concept art, and brands can produce logos; the possibilities are endless. People will no longer need expensive equipment, art supplies, or be trained in drawing fundamentals to be able to have art.
With everything, there are also negatives. One of them being, the issues with artists and copyright. As AI-generated images increase in quality and accessibility, it may put many artists out of work. Furthermore, AI-generated images can easily replicate the style of artists simply by putting their names into prompts. Many artists may feel that their work is unjustly copied and want to opt out. Moreover, there is the issue of copyright; who is the artist of AI-generated images? Is it the user who imputed the text? Is it the AI? If the text includes a specific artist, should the work belong to them? All these issues must be dealt with to ensure the fair use of these images.
Prompt: a builder [Photo Credit: OpenAI]
Prompt: a flight attendant [Photo Credit: OpenAI]
Next, DALL-E itself has its own biases. For example, if “builder” is imputed, the AI includes only images of white men. When the “flight attendant” is submitted, only Asian women are pictured. Another complication is the misuse of the program to create explicit content. DALL-E has blocked explicit words, but people can bypass this by creating a prompt that is visually similar to something the filter blocks. For example, instead of using the word “blood,” people can use “red liquid” to get the same effect.
Although DALL-E and other AI-generated image programs are incredible, creating limitless possibilities, it also has their downsides. There are issues surrounding copyright and artist, DALL-E’s biases, and the misuse of AI. People should be impressed by the astounding outcomes of AI-generated images and cautiously optimistic about its future.
Although DALL-E is not widely available for public use, DALL-E mini, now rebranded as Craiyon, is available and can be used here: https://www.craiyon.com/
References
Openai. “DALL·E 2 Preview - Risks and Limitations.” GitHub, 19 July 2022,
https://github.com/openai/dalle-2-preview/blob/main/system-card.md#bias-and-represent
OpenAI. “Dall·E 2.” OpenAI, OpenAI, 14 Apr. 2022, https://openai.com/dall-e-2/.
Romero, Alberto. “DALL·E 2, Explained: The Promise and Limitations of a Revolutionary AI.”
Medium, Towards Data Science, 16 June 2022,
https://towardsdatascience.com/dall-e-2-explained-the-promise-and-limitations-of-a-revol
Vox, director. The AI That Creates Any Picture You Want, Explained. YouTube, YouTube, 1 June
2022, https://www.youtube.com/watch?v=SVcsDDABEkM. Accessed 2 Aug. 2022.