DALL·E 2 Review and Demo
Open AI’s Dall-E 2 is now available to everyone. What can it do, what can’t it do, and where is it going?
Introduction
DALL·E 2 is the latest text-to-image AI from Open AI that generates images based on text prompts. Sometimes they are really great, at other times primitive, and sometimes just broken.
It is available as a website and API that requires an account to use. The project is closed-source and is only available through the website (or API)
DALL·E 2 was created by training on private and publicly available images. Many of these images are from the Internet; if the subject is popular on the Internet, it is likely that it will generated good results. For example, DALL·E 2 really knows what a Shiba Inu looks like.
Accessing DALL·E 2
To access DALL·E 2, sign up for it at:
https://openai.com/dall-e-2/
You will get 50 free credits your first month. 15 free credits will replenish every month after that and Free credits expire one month after they were granted. You can also buy credits, but note that all credits expire 12 months from the date of purchase.
How it works
DALL·E 2 takes a prompt and generates 4 images. The more detailed or descriptive the prompt, the more likely that an interesting output will be generated. You can use a generated image as a basis for another generation.
There is a history of your generated images on the right side of the screen, so your previous work is not lost.
The images generated by the prompt will not always be the same: you can generate several images (or generations of images) from the same prompt.
The images can be grouped in a collection, and images and collections can be made private or public.
What is a Prompt?
The prompt is the most important part of the ‘creative process’. The biggest challenge is crafting a prompt that will result in a desired or interesting output.
DALL·E 2 is really good at mainstream or very well-known artists’ styles. It is also effective when prompts describe photographic lenses, film, lighting or other compositional details.
Perhaps unsurprisingly, given its training with images from the Internet, any popular images or memes should also be well understood.
Demos
On to the demos! These are prompts that have created interesting results, I have tried to credit where I found them.
Links to the high resolution public collections included.
Andy Warhol Animals
Prompt:
A painting of a XXXX in the style of andy warhol
Full collection (Credit)
2d Sports Team Logos
Prompt:
2d ferocious XXXX head, vector illustration, angry eyes, football team emblem
Full collection
Animals in Turtlenecks
Prompt:
happy XXXX wearing a yellow turtleneck, studio, portrait, facing camera, studio, dark bg
Full collection (Credit)
What can’t it do?
DALL·E 2 can’t spell. Seriously. Words are just junk (at the moment anyway). You need to be very prescriptive if you want a phrase to appear in your output. Even then, good luck.
It has some problems with eyes, extra legs and arms. It generally knows what things should look like, but those ‘things’ are based on images, not rules (ie people have two arms and two legs). To DALL·E 2 those aren’t rules, but suggestions open to creative interpretation.
Open AI has built in some content protection that prevents gore and adult content, but also blocks (IMHO) more innocuous prompts.
Examples of prompts that will generate a content warning are:
- “I poop sparkles” <unicorn emoji> sticker illustration
- An illustration of Xi Jinping as Winnie the pooh
- A studio picture of Vladimir Putin in drag with lip stick and black background 4k
Conclusions
DALL·E 2 is worth trying out. These prompts have been selected because they create interesting results. Crafting good prompts can be a bit of a challenge and the underlying DALL·E 2 AI is also changing.
How should DALL·E 2 be used? As a tool to brainstorm, explore ideas and create starting points for further development.
Is DALL·E 2 a threat to artists? Just like Photoshop or Illustrator didn’t replace conventional artists, DALL·E 2 will augment the creative process and be another tool.
What is next? The use of DALL·E 2 to outpaint existing images is interesting. In brief, a starter image is provided in addition to the text prompt and the starter image is incorporated in the generation.
Recommendation: Sign up for DALL·E 2 and spend a few minutes exploring. After all, it has a low barrier to entry.