DALL-E is spectacular. I’ve been using it to make everything from abstract inspiration for acrylic paintings to my favorite Nintendo characters reimagined as renaissance era royalty. But creating great AI art takes more than a button press: you also need a well-engineered prompt. In this article, I’ll give my advice to getting better results, having created a few thousand AI-generated images. I’ll also talk through some of problems with AI art and potential mitigating factors.

How to write a great DALL-E prompt

The two big ones:

  1. Describe the specific content you want to see; include details about the foreground or background
  2. Add what medium you’d like represented, such as a photo, pencil drawing, watercolour painting, digital art, etc. My favorite is “oil painting”

Additional style tags:

  1. Artist (if you want an impressionist oil painting, write Monet or Cezanne; if you want something freaky, write Hieronymus Bosch) a. You can also name an artist’s style, such as “cubism” or “surrealism”. But IMO artist names work better, and allow more precision
  2. Lighting - brightly lit, gloomy
  3. Color scheme - e.g. triadic, complementary, sepia
  4. Composition - landscape, close-up, ultra wide-angle

Here’s a few samples of artwork generated using the features above.

Emperor fox in red robes on the throne with a gold crown, rembrant oil painting, psychedelic color scheme 404 image not available

An oil painting of a skeleton riding a clydesdale in the style of dali 404 image not available

Leochares Belvedere Apollo holding a boquet of flowers impressionism 404 image not available

A 3d model samus aran relaxing by the pool, pink and blue color scheme, with a lavish background 404 image not available

Digital art Tony hawk hoodrats rustic family dramatic lighting cyberpunk Hieronymus Bosch 404 image not available

Oil painting of Marble veins triadic technicolor color scheme 404 image not available

Problems with AI art

  1. Copyrights - the works AI algorithms trained on are often copyrighted, and artists are currently getting scammed by having their work stolen for training data.
    1. This is a major problem in the algorithms’ current form, especially as AI art becomes better monetized. I expect we’ll see a lot of change in the industry, as artists restrict the use of their work from AI training. As a result, AI-generated art could become more expensive to produce, since more work will be required to source diverse training data.
  2. Death of art - as more art becomes AI-generated, there’ll be less novel art produced, which could create a self-destructive spiral.
    1. I don’t believe this will actually manifest, outside of the typical undulations of capitalism seen from the industrial revolution to the advent of the internet. Novel artwork will still be incentivized, especially in high culture, and AI art will continue to have weaknesses, such as producing large scenes with complex interactions. Additionally, we’ll see the rise of AI-assisted artists, which will produce high volumes of quality content. Like with previous technological innovations, some artists will be displaced, while others will learn new skills and thrive. Personally, I believe UBI is the best way for Americans to care for those displaced by industry changes, but how and why are outside the scope of this article.
  3. Bias in the algorithm - the AI will be at least as biased as its training data, and may reinforce harmful stereotypes.
    1. TBD if AI art can overcome bias in training data; even self-driving cars are affected by this. Currently projects like DALL-E employ swaths of personnel to screen the algorithm for bias and flag & restrict its problematic output. Continuous investment will be necessary.
  4. Hate speech via art - AI can be used to create harmful and easily disseminatable images.
    1. This is an unfortunate reality, but nothing new. The written word, for example, is used in most online hate speech, but its benefits far outweigh the costs. Considering that the alt-right already has a deep well of established memes, I don’t expect AI to produce a step change in their potency. One execption is AI depiction of real people in fake scenarios, similar to the deepfake videos starting to be produced. Hopefully algorithms to detect fake imagery will keep up with the algorithms to produce it. And hopefully we’ll see improvements in online content moderation. Exactly what those could be is worth another article, or many tweets, posts, and books.