It is the first artificial intelligence (AI) capable of generating convincing images that is both easy to use and free. DALL-E 3, provided by OpenAI and offered by Microsoft under the name Image Builder, has been available since October 3 to people with a Microsoft account.
Generating an image requires around twenty seconds of waiting. A time that increases exponentially when we exceed a threshold of fifteen images per week. To help you understand the usefulness – or not – of this astonishing tool, here is a summary of what we can do, step by step, but also of the images that it will refuse to generate.
What is possible
• Imagine any type of character. You can ask the AI to create an animal, a robot, a human, or even a chimera combining all of these at once. To do this, simply describe this character in clear French, using precise words. This is what we call writing a prompt, which can be translated as “instruction.” For example: “a slightly fat black alley cat.”
If you wish, you can choose the character’s posture, then define a framing (wide shot, tight shot, etc.). For example: “chest shot of an elegant 75-year-old lady, with bright and determined eyes, wearing a geometric dress.”
• Imagine any type of decor. You can choose to place your image in an existing city, or in the landscapes of a particular region, as long as they have a certain notoriety. For example: “the snowy Oslo Opera House seen from the icy sea.” However, it will probably take several attempts to get an image close to what you imagined.
Alternatively, you can imagine a completely fantasy setting. Like this one: “in the foreground, a star-shaped swimming pool with an off-white background, protected by a glass bubble, in the background a view of space with a sun occupying the top of the image” .
Note, however, that your imagination will occasionally be limited by DALL-E 3, which tends to refuse certain prompts for inscrutable reasons. We were, for example, only allowed to generate the following image after replacing the word “rock” with “rock”.
Alternatively, you may prefer an interior decor. And if you wish, invite a character created earlier: “a slightly fat black alley cat in the Mont Blanc ice cave”.
• Choose a style. Start by choosing an art form: photography, painting, or comics. You can be more specific and cite an artistic movement, or even a particular artist, preferably iconic, otherwise the result will rarely be convincing. A famous Dutch painter, for example. Note, however, that the AI refuses to pastiche certain contemporary artists.
You can be even more specific by choosing an artistic technique. For example, for a photo, you can ask that the image respects the aesthetics of a medium format square film camera and specify that the lights are placed on the sides of the subject, to highlight the reliefs:
• Imagine an improbable scenario. DALL-E 3 is not intended to produce only realistic images. You can use it to create scenes that are grotesque, absurd, or contrary to the laws of physics: let your imagination speak.
What doesn’t work
• Feature a contemporary celebrity. The AI refuses to create images that contain them. We can try to get closer to an actor’s physique by describing it very precisely, but the results are generally disappointing. On the other hand, we can often summon a historical character, such as Salvador Dali, in an image and stage him in a fanciful way, by making him dance with death for example. The resemblance to the character is striking, but this is far from always being the case.
• Create sexually explicit images. Microsoft Image Creator refuses to generate images depicting sexual acts. And while it occasionally accepts requests for naked characters, it ultimately generates them with clothes. Internet users have undertaken to list the specific words banned by Microsoft, but these appear to vary over time. Attempting to circumvent these limitations may impact your account as a threatening message is displayed for each request deemed problematic: “Further policy violations may result in your access being automatically suspended. »
• Create violent images. In its content policy page, Microsoft places particular emphasis on images glorifying suicide, graphic violence and macabre elements, as well as visuals evoking terrorism or violent extremism. “We prohibit the use of Image Builder to produce content that advocates violence against others,” Microsoft says. In addition, the Image Creator refuses to generate certain settings, such as the American base at Guantanamo.
• Transform an existing image. Unlike some competing AI, Image Creator is not capable of importing an image to transform it according to our instructions. In addition, it is not capable of describing existing graphic content by summarizing it in a prompt, which competing AIs allow. However, this function allows us to better understand the way in which AI reads images as well as to regenerate visuals close to the original, by adding variations.