Ultimate guide to DALL·E 2: how to use it & how to get access

Amazing, isn’t it? Become an AI artist and discover everything that DALL·E is capable of in this definitive guide.

🎨 How to use DALL•E

What is DALL·E?

DALL·E is an AI art web app, designed by Open AI, which uses artificial intelligence to turn sentences (like ‘A grey horse galloping along a beach at sunset’) into images. It can also accept images uploads and modify them.

Is DALL•E real?

Yes, my sweet summer child – I can I assure you it is very real!

How do you use DALL•E?

Using DALL·E to create AI artwork is remarkably simple.

There are three main ways to create an image: text prompts, variations, and edits.

Let’s take the most well-known method first – text prompts.

How do you prompt DALL•E with text?

The main interface for DALL·E is this: literally just a text box.

Click ‘generate’, and wait about twenty seconds.

DALL·E presents you with ten six images based on your prompt.

At this point, you can click on any image to see it at full size (1024px x 1024px square), download it as a PNG, save it to your personal collection, or create a shareable link. It really is that simple.

That’s it! If you like, you can tweak the wording of your description to get a different type of result, or just ask for something completely different. (A word of warning: users are currently limited to 50 requests in a rolling 24 hour window!)

Apart from prompting with a sentence, you can also prompt DALL·E with an image.

There are two ways. to do this: a variation or an edit.

How do DALL·E variations work?

A variation simply prompts DALL·E with an image, rather than written text.

In response to the image provided, DALL·E creates nine five additional images, that it thinks are pretty similar.

Wait twenty seconds, and DALL·E provides five fresh options.

Notice that our latest set of astronauts tend to share the upload’s regal, religious art style, with a preponderance of mournful expressions. They all have a similar layout, too.

Variations: creating new images based on existing artwork

It’s not just DALL·E’s own outputs you can vary: you can upload any image, so long as you have permission to use it.

For example, you could upload an old painting in the public domain, like the Mona Lisa, or another free-to-use work, like a Creative Commons photograph.

In the flow below, we’ve started with an image from Unsplash by Sharan Pagadala, then used variations to generate ten similar pictures.

Notice how the overall colours are the same, but are being deployed in different areas of the image – in some, for example, the sky is starting to take on the hue of the suit. In some variations, the character remains within the cloud, but in our winning version, the cloud has become a sort of sofa.

As with our Austenaut, we could continue to make variations-of-variations, until we ended up with somewhere altogether different, depending which of the 6 images we chose each time.

Creating a variation of your own work

If you’re an artist, you can also upload your own work, and create variations of that. For instance, the character artist THIP uploaded a character sketch he’d created, of a helpful robot with the body of an arcade machine, on the top-left.

It’s a quick and easy way to create a bunch of similar-feeling characters – this could save time in crowd scenes, for example. But bear in mind that, currently, OpenAI claims the copyright to the variations generated.

How do I add prompt instructions to a variation?

You can’t.

For instance, we can’t upload Astronaut Austen to DALL·E and ask it to make her look happier, or as if Picasso had painted her, or anything else that controls the result.

Essentially, DALL·E creates its own internal ‘description’ of the source image (admittedly using strange alien adjectives like “A massive [ENTITY 921298ASA], very [QUALITY 91283202AAB] with major overtones of [ASPECT 2839HSJ3]”), then gives that description to itself as a prompt, then generates nine more images based on that.

Fortunately, we CAN control how DALL·E changes images with ‘edits.’

What are edits in DALL•E?

Edits are the third and final way to prompt DALL·E – like variations, we provide an image, but this time we can control it by giving DALL·E a blank space to amend.

Once again, DALL·E makes it super-easy to do this with an output – just hit edit and use the eraser tool to get rid of whatever you don’t want to keep.

Unlike variations though, we can now add a description of the overall new image we’re trying to build. Let’s put Astronaut Austen out into space, where she belongs, by describing the scene:

‘Astronaut Jane Austen, staring into the cosmos, floating through deep space, beautiful galaxies in background.’

And away we go!

Just like variations, we can use ‘edits’ on any image we’re permitted to use.

Below, we take a photo on Unsplash by Jonathan Formento, erase the central area, and ask for ‘intricate alien ruins of crumbling black stone, in an extraterrestrial landscape’.

Notice how DALL·E leaves the un-erased parts of the image completely untouched. We couldn’t have asked for a ‘whole-image’ shift like ‘on a stormy day’, because all the sunny parts of the image would still be there.

Creating edits of your own work

Your can erase portions of your own images, too.

What is out-painting?

Unlike ‘in-painting’, where we use the Edit tool to delete interior parts of an image then get DALL·E to fill it in then blanks, we can instead get DALL·E to reveal ‘more’ of an image.

We take our original image and bring it into Photoshop (or similar) then shrink it down, leaving transparent space around it.

Then we export this as PNG and then upload our new file to DALL·E.

↔️ A practical use for this tool this tool is to quickly improve DALL·E outputs where the framing is too tight, and parts of the subject have been cropped out.

How can I create landscape/portrait photos in DALL•E?

DALL·E only generates square outputs, but by using out-painting, you can stitch together wider images!

Simply download an output, bring it into a photo editing tool, and drag it to one side of the canvas, leaving the gap you create transparent.

Save this file as a PNG and it to DALL·E, choose ‘edit’, and prompt for what you want to appear in the gap (we asked for “alien spacecraft”) – then pick a variation.

Now you can simply align this second image with your original in Photoshop, to create a wider scene!

There’s no limit to how far you can take this – epic murals, panoramas and even sky-scraping vertical images are within your grasp.

In the images below, we’ve used out-painting to create portrait, ‘Instagram Story’-sized images. Because we can describe what to put in the blanks, we can experiment with describing the characters, then create a separate description for the alien shape above.

It’s even possible to create Very Big Squares, expanding in both directions at once – but you’ll need to be methodical, so everything lines up!

What is prompt engineering?

‘Prompt engineering’ is the art of describing an output in such a way that DALL·E delivers the result you want.

Because the magic text box is the main way of controlling DALL·E, the terms we use to describe each request (or ‘engineer a prompt’) are paramount.

Compare the evolution of these two images from Daniel BLN:

(See the full breakdown of optimising this image)

DALL·E has seen over 650 million images during training, but has never been explicitly programmed to create images one way or another – it didn’t get ‘lessons’ in brushstrokes or camera film – so not even DALL-E’s inventors at OpenAI know precisely what DALL·E understands. Yet.

First, as a user, knowledge is power. For instance, you can ask DALL·E for a photograph of something: but you can also specify the type of camera, film, lens, lighting conditions, and so on.

As a non-photographer, you might not be familiar with how technical choices contribute to a photograph’s ‘look’ – so even if you’re only working in AI, you might find yourself suddenly needing to swot up on photography.

Read more: 17 creative photography styles you can achieve with DALL•E 👉

But that’s only half the battle: we also need to collectively discover what DALL·E understands. Some prompts, it can execute with absolute conviction, others it struggles with. Together, resources (like this website!) will start to emerge, documenting how well real-world DALL·E responds to different styles of art – what descriptions get results, and which are better rephrased.

🔑 Getting access to DALL•E

Can anyone use DALL•E?

Not yet – access to DALL·E is currently invitation only.

How do I get an invitation to use DALL•E?

You can sign up to the waitlist on OpenAI’s website. There is also a separate ‘artists and creators’ track, which operates on its own schedule, where OpenAI are inviting creatives to test out the platform.

How many people have DALL•E access?

30,109, as of June 23, 2022.

How many people are in the waitlist queue for DALL•E?

Over 100,000 people joined the waitlist in the first 3 days. It’s not known how many have joined since.

How quickly is OpenAI adding new users?

As of May 18, 2022, new users are being added at the rate of about 1,000 a week. Assumedly, this will start to speed up – I’d hope there’s not going to be a two year wait to get in!

How do I get creator access to DALL•E?

There is no official application process; you will need to find a way for the right people at OpenAI to know who you are. Be cunning, like a fox!

Current participants are from a range of backgrounds and with different artistic specialities. You do not necessarily need to be ‘famous’, experienced, or an influencer.

Some things that I suspect might help are: a previous or ongoing project involving AI creation, a track record of sharing art or documenting your artistic process publicly, a clear proposal for a project or experiment you would use DALL·E for that’s easy to visualise.

Things that I suspect will be unhelpful: an interest on pushing the ethical/content boundaries of DALL·E, ideas involving NFTs, extremely abstract academic proposals, and just telling them how desperately you want access.

How much does DALL•E cost?

It’s free to use DALL·E for now!

But expect that to change: every prompt DALL·E handles costs OpenAI a bit of money (it’s computationally ‘intense’, in their words) so they’ll want to recoup that.

Is DALL•E open source?

No.

What kind of computer or hardware to I need to run DALL•E? Can I download it?

DALL·E is a web app that runs in your browser – the image generation happens on OpenAI’s servers. The good news is, then, that you can use DALL·E on any device, but the bad news is you can’t download the code and tinker with it.

Are there limits on DALL•E usage?

Yes – currently you can only prompt DALL·E 50 times in any 23.5 hour window, giving you a total of 500 images. This limit includes written prompts, variation requests and edits – any action that gets DALL·E to generate six new images.

There are no other settings – you can’t change image resolution, canvas size or shape, or anything to do with the algorithm used.

All images are 1024 x 1024 pixels square.

There are also content policy rules, described below.

👨🏽‍⚖️ Rules, limits and legal stuff

What are the rules of using DALL-E?

DALL·E’s creator, Open AI, has a clear content policy. Outputs should be G-rated, while images around politics and controversial issues are also forbidden. Meanwhile, using an ‘innocuous’ image for socially disruptive purposes, like fake news or harassment, is also against the rules.

Since June 21, you aren’t allowed to create prompts using public figures (e.g: ‘Ellen DeGeneres teaching math to a duck.’)

You can’t upload images to vary or edit that you don’t have permission to use, e.g: from another artist.

You can’t upload any images of human faces – even your own.

What safety measures does DALL•E have?

📃 A fairly conservative content policy that limits even mildly controversial material

✋ OpenAI closely controls access, so they can ban users at their discretion

🙊 Certain words and phrases are blacklisted from prompts altogether

🙊 Prompts that trigger ‘blocks’ are logged for further review, and deliberate attempts to create clearly out-of-policy images result in account closure

🛑 Prompts with the names of public figures are blocked

🛑 Image uploads with human faces are blocked

🔎 The image outputs are themselves examined by AI for inappropriateness, before they’re displayed

🙈 The image set DALL•E was trained on was filtered for adult material; DALL·E can’t recreate what it hasn’t seen

🚩 A specialist ‘red team’ was tasked with finding ways to circumvent DALL·E’s safety features before launch, so loopholes and exploits could be fixed

⏳ OpenAI is gradually onboarding users step-by-step, can see what % of users attempt to contravene the content policy and adjust accordingly; they can slowly gradually expand ‘generations-per-week’ to track what percentage of outputs are ‘accidentally offensive’ despite an innocuous output

Who owns the intellectual property and copyright of DALL•E images?

Currently, OpenAI claim copyright over all images created with DALL·E.

You also have to credit DALL·E with the work, not pass it off as your own. Usually you can do this simply by keeping the ‘coloured squares’ watermark in the bottom-right corner which they describe as ‘DALL-E’s signature.’

In practice, OpenAI are not defending their copyright for personal use – you are relatively free to share these images online, in accordance with the sharing policy.

OpenAI recognise the current policy is artist-unfriendly, and is a placeholder during the test phase. Expect a more nuanced copyright policy in the near future.

One interesting twist: in fact US law currently holds that a computer output cannot be protected by copyright, so as precedent stands DALL-E’s images can’t be owned by DALL·E – or you. Ever!

Transformative works – e.g: creating a sculpture of a sculpture you got DALL·E to design – are permitted, and these works then belong to you, not OpenAI.

Are there any other copyright concerns?

There are other intellectual property issues you should consider – largely, these are the exact same as if you were creating an image yourself in Photoshop – with a few nuances. They will be more relevant when

• Copyright in artwork: artists can’t copyright a ‘style’, so you can happily generate images ‘in the style of X.’ But if you create an image based on a specific famous image (that DALL•E knows about) created by an artist who is still alive or died after 1952, it could be challenged as infringing, even if your output is quite different. The evidence that your prompt deliberately referenced the specific work might also be incriminating!
• Trademark and character copyright: For instance, if you generate an image of Pikachu eating a pizza, and try to use it advertise your pizza restaurant, Nintendo might litigate because they own the copyright of the character itself, and the character is also a trademark they own.
• Personality rights: If you create an image of a celebrity doing something (like Beyoncé drinking your nootropic soda) and try to use it commercially, their ‘personality rights’ also come into play.
• Defamation: If you created an image of a real person doing something that makes them look bad, they could litigate, especially if it’s convincing. (This is against OpenAI’s policy anyway.)

Yes, it’s true that DALL·E creates every image ‘from scratch’ – the outputs aren’t crude combinations of specific images.

But a common defence against copyright infringment is having never seen the original work – that you created something similar by coincidence. Because DALL·E is trained on 650 million images, and users have no way of knowing what they are, it would be hard to demonstrate this.

In short, DALL·E is a whole new kind of tool with no good precedent; the next decade will be sure to offer some interesting test cases.

Remember, ‘being right’ is not enough: you have to be able to defend yourself. In 2010, the photographer Jay Maisel sued an innocuous hobbyist for creating a pixel art version of the album cover for Miles Davis’ ‘Kind of Blue’, demanding $175,000. Although said hobbyist believed this was clearly fair use, he could not afford the legal fees to defend the claim, or the risk of losing altogether, so had to pay the photographer $32,500 to settle.

In general, it’s not great to get sued by an artist or their estate.

What about the images DALL•E was trained on? Isn’t DALL•E copying them?

DALL·E was trained on hundreds of millions of images.

Some have contended that this is improper, and that consent from those whose works were included ought to have been sought. (The actual images used are not publicly known.) Others claim this is an outright infringement of the artist’s IP.

In practice, it’s a bit more complicated. DALL·E is only basing outputs on the training data in a very general sense.

• trainee artists at art school study thousands of paintings to learn their craft;
• if you wanted to paint a Parisian street scene, you might first look up a lot of photos of Parisian streets on Google Images;
• similarly, DALL•E has looked at a lot of images to learn what the world looks like, or what particular qualities makes an image ‘cute’, ‘stormy’, ‘impressionist painting’ or ‘Polaroid’

Just as the trainee artist doesn’t need to pay a special fee to every prior artist who’s work has shown them the principles of composition, it’s not clear that DALL·E owes specific compensation to the creators in the training set – it may stand on the shoulders of millions of giants, but so do we all.

Secondly, the training images aren’t ‘stored’ within DALL·E – the generator isn’t ‘looking up Image 92,291,92 and Image 371,023,541 and then combining them’ to build an output. Instead, DALL·E just remembers the general principles it learned from looking at them. (See how DALL•E creates images.)

It’s the same method you or I might use to draw a dog – we draw upon our many memories of seeing dogs in general, which have in turn created a coherent mental concept of ‘what makes a dog look dog-like’ and ‘common poses dogs are in when pictured.’ When we do so, we aren’t consciously referring to any specific image or canine encounter – it’s all of them at once, and none of them in particular.

Can I sell DALL•E images? Can I use them in my job?

During the ‘test phase’, all DALL·E outputs are currently under a ‘no commercial use’ policy.

Specifically, this means you can’t use DALL·E images to complete commissions on Fiverr or Upwork, sell them on Shutterstock, or bluntly ‘sell access to others’ in a “send me $5 and I’ll run your prompt” manner. Based on a briefing, it seems like it’s this direct monetisation is the target of this policy.

Many other commercial uses could be construed as forbidden (e.g: using DALL·E to brainstorm ideas in a business meeting), but in a briefing OpenAI encouraged creators to incorporate DALL·E into their creative practice, so it’s a gray area.

If the final project is given away for free, like as a cover for a downloadable album, it’s fine, but then if that same album appears on Spotify and the artist makes $0.000008 from it, perhaps not.

As mentioned before, if you transform an output into something new (like a product design) the new output is not subject to these restrictions.

👀 Behind the scenes

Who created DALL•E 2?

DALL·E 2 was created by OpenAI, an organisation in San Francisco. Credited in the academic paper explaining the principles behind the tool are Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu and Mark Chen.

About 300 people work for OpenAI. In 2019, Microsoft invested $1bn in the company. Elon Musk is a donor and used to be on the board. Their mission is to ‘ensure that artificial general intelligence benefits all of humanity.’

OpenAI’s other products include the writing & text AI, GPT-3, and Jukebox, an experimental music generation system.

Is DALL•E 1 available for use?

No, DALL·E 1, which was revealed in 2021, was never released for public use.

How does DALL•E actually work?

The underlying technology is unsurprisingly complex, but basically comes in two parts.

In the training phase, System 1 (called CLIP) was shown millions of images with text captions, and through this process, gradually learned to associate particular words with particular visual attributes.

When you prompt DALL·E, CLIP turns your sentence into a matching set of ‘image attributes’ that describe the ‘gist’ of what the output should look like.

Then, a related model, called unCLIP, is given ten images of pure noise, and told that these are extremely corrupted versions of an image with CLIP’s attributes.

unCLIP tries to ‘fix’ the image by modifying pixels, so that at each step the image has more of the attributes CLIP has provided – an image that ranks more highly for ‘astronautiness’ and ‘Jane Austenness’ for example.

It carries out this same process whether you’ve requested a photo, painting, or anything else – it’s all the same to DALL·E.

DALL·E doesn’t ‘paint’, or set up ‘objects’ with ‘light sources’ and camera lenses – every image just slowly emerges from the mist. Like magic!

To be clear, DALL·E also doesn’t combine specific images it’s been trained on, like a Photoshop whiz. It’s just… brand new.

For more information, read:

• A fairly straightforward explanation from Adityara Ramesh, inventor of DALL-E
• A deeper dive, with diagrams, from Assembly AI
• The original academic paper, on arxiv

What images were used to train DALL•E 2 ?

The CLIP and DALL·E datasets, which were both used to train the system, contain 650 million images combined. ‘Adult’ images were removed from the data prior to training. The specific images or libraries used have not been publicly revealed.

How has DALL·E 2 (and this article) changed?

June 21, 2022
Previously, OpenAI prohibited sharing images of human faces – this prohibition is now removed!
Previously, you could ask DALL·E to create images of celebrities (e.g: ‘Elon Musk riding a unicycle.’) This is now against the content policy.
Previously, you could upload images of human faces for editing, out-painting or variations. Now uploads with human faces in (even your own!) are blocked.

mid-June 2022
OpenAI quietly released a model update improving DALL·E 2’s ability to generate realistic human faces.

June 6, 2022
Open AI reduced the number of images generated per prompt to 6. (Previously, it was 10.)