A prompt designates an instruction, which is sent to an artificial intelligence (AI) algorithm specializing in the generation of content – such as text or an image.
What is a prompt?
A prompt is an English word that designates any written command sent to an “artificial intelligence” specialized in the generation of content, such as text or images. In short, it is an instruction that is given to an algorithm, which will then interpret it and propose a result. The prompt can also be thought of as a request or an order.
What can be generated with a prompt?
It all depends on the specialty of the generating AI. In the case of ChatGPT, it is text. We ask him a question and the chatbot responds in writing. There follows an exchange, as if it were a conversation in a messenger with one of his contacts. It can also be images (DALL-E, Midjourney) or sounds and music (MusicLM).
Today, the tools are specialized in one area, but we can expect the imminent arrival of “universal” platforms, that is to say, capable of generating results in different forms (text, image, video , her…). We could also see a multimodal approach appear, by offering an instruction that would combine writing and an image.
How to generate a prompt?
On DALL-E or ChatGPT, simply enter your text in the area provided for this purpose. In the case of Midjourney, you must type a specific command, which begins with “/imagine” (without the quotes), followed by your description. For example, “/imagine a glowing red panda crossing the Milky Way while sitting on a star.”
Note that there are also other commands to launch a prompt on Midjourney. One of them is called “/blend”. This command is used to merge several images on Midjourney — it’s up to you to provide them. Beware, on this subject, of two things: if you use copyrighted images, the artists may object to it. Also, the images you send are public.
How to make a good prompt?
There is no secret to making a good prompt. You have to be imaginative and provide as much material as possible for artificial intelligence. The more precise and detailed your description will be, the better the result will be. This means that you will have to add keywords in your text so that the algorithm has everything that interests you.
You can also go quite far in your quest and obtain quite spectacular results. Giving instructions on lighting, style, viewing angle. By adjusting the color or mood. By playing on the zoom or the composition. Sometimes you won’t get anything good on the first result. Then, you will have to modify your prompt.
There are online resources that can help you. For example, this site has written a PDF to serve as a compendium for DALL-E prompts. OpenAI, the company that developed DALL-E and ChatGPT, provides on this other page a guide to generate text with its chatbot. But, the best is probably to experiment to see how things work.
There remains one question: should the instruction take place in English or in French? It was found that the instructions were understood in French. However, as these are tools designed by American companies, the use of English may be an option to consider. If you struggle to have all the adequate vocabulary for your prompt, don’t hesitate to use DeepL.
Some prompt examples
Here are some prompts passed on DALL-E:
- A portrait of a dog in a library, Sigma 85mm f/1.4 (result)
- A gray kitten standing on a pizza in outer space. The kitten is eating a piece of pizza. Pizza slices flying with angel wings in background, dark cyan galaxy and stars in background, 4K photoshopped image, look at that detail (result)
- a male American high school student reading a newspaper, in Chinese watercolor, award-winning painting (result)
- full body photo of a horse in a space suit (result)
- Jackson Pollock, Air Jordan sneakers, digital art, product photography (result)
There are many other examples to discover on Twitter, Instagram and the rest of the web, whether with DALL-E or Midjourney. What is meant to show here is the variety of renderings that can be achieved whether your prompt is brief or detailed. It all depends on the degree of orientation you want to give to the AI, by mobilizing vocabulary from art or photography.
Where can I see other prompts (and results)?
If you lack inspiration or want to see other creations, there are online tools and galleries. There are also lots of visuals to find on social networks.
This site generates prompts on the fly (like “Darth Vader with gray hair holding bushes in a football stadium, Realistic painting, 4K HD” or “Cucumbers glowing and vibrating in a bus stop, Cover of Rolling Stone magazine”). It is then up to you to generate them! On DALL-E, you can also re-generate alternative images on already used prompts.
In the case of Midjourney, you can see the creation of other Internet users. For this, you need to join the Midjourney Discord. You can consult images already created, but also recover them if you wish. You will have the choice to request other variations on visuals that have caught your attention, or to relaunch a whole new generation.
How to make better prompts?
Your best asset to improve your prompts will be to use precise vocabulary to write the request and to make a description as complete as possible. There is also a technical aspect, perhaps less known: these are the parameters. This is not true for all content-generating AIs. Midjourney, for example, uses it, but not DALL-E.
Among the classic commands, there are those that set the size of the displayed image. For example, adding “–ar 4:3” (without the quotes) calibrates the visual to a TV aspect ratio display. There are other technical commands, such as “–chaos” for the level of abstraction. This site lists several parameters that you can add to your prompts.
If you liked this article, you will like the following ones: do not miss them by subscribing to Numerama on Google News.