If you need images for your social media campaign or website but you can’t find what you want on stock image sites, then generative artificial intelligence can help. With a little practice, you can get an AI to produce all kinds of images for you.

If you have been following this series, then you will probably have experimented with large language models (LLMs) such as ChatGPT. As with LLMs, there are a variety of image-generating AI systems to choose from. The highest profile are DALL-E, Stable Diffusion and Midjourney. These are available not just through their own websites and apps but are often embedded in other services too.

The big three

DALL-E, for example, is the underlying AI model for Microsoft’s Image Creator tools. That isn’t surprising because Microsoft is the largest investor in OpenAI, which is the company behind DALL-E and ChatGPT.

Stable Diffusion, meanwhile, has been released as open-source software by its makers Stability AI, which means there are lots of tools available that use it. You can even download it and run it from your PC. Overall, the best way to get started is with Stability AI’s own tool, Dreamstudio. Finally, there’s Midjourney, which is currently only available through the chat service Discord. More on that later.

Aside from the ‘big three’, various other companies have added image generation tools. Google has ImageFX, which uses an innovative series of dropdown menus to help refine your prompt. Adobe, known for photo editing software Photoshop, has added an AI tool called Firefly. And stock image companies, Getty Images and Shutterstock, have added their own tools. Which one you end up using will depend a lot on personal preference.

Prompt engineering for image generators

Prompt-writing for image generators is like prompting LLMs: explain in natural language what you want to produce, then give it a moment to work on a solution. Creating an image takes longer than producing text, partly because LLMs share the start of their answer before they’ve written the end, whereas an image generator must complete the task before showing you anything.

LLMs also have multiple output types. They can answer questions, generate new writing, summarize documents, and so on. Depending on what you want to achieve, you might write your prompt in different ways. In a sense an image generator is a simpler proposition because it just produces images. You just have to provide a good description.

However, there are variables within that seemingly simple task. You should begin by describing the image content: the kind of scene, what it contains and what action - if any - is taking place. Then move on to how the image is portrayed, such as mood, lighting, framing, and so on. The mood could be “inspiring” or “ominous”, for example. Lighting could be specific (“sunlit”) or more general (“moody”). Finally, you can add specific details, for example, “a woman stands on the right of the frame, looking towards the camera”.

Challenges with prompting

Image generators don’t like long prompts, though. AI gurus often recommend multi-paragraph ‘mega-prompts’ for LLMs, but those should be avoided here. There is no hard limit but Zapier suggests a 60-word maximum for Midjourney and 380 characters for Stable Diffusion. The longer the prompt, the more likely it is that the AI will ignore words at the end. This will probably improve but keep those limits in mind for now.

Finally, if you have used LLMs regularly, then you will have become used to the conversational interface. You can ask a question, get a response, ask for more detail, and so on. If you are creating text, then you can paste responses together to get the best combination. This doesn’t work for images because if you ask for changes, the AI will generate the entire image again. That means it might fix the bits you didn’t like but also scrap the ones you did.

Even if you specify an area that you don’t want to change, the AI might ignore you. In DALL-E, for example, you can shade an area and give the AI an instruction about it. I had it produce a picture of a sports car driving through a mountain range, then shaded the car and told the AI I wanted to keep the car and change the background. The AI changed the background as I asked but re-drew the car too.

Tips and tricks

1. Create lots of images. Professional photographers don’t just take one photo and assume that will do. They take lots and pick the best. The people showing amazing AI images online are doing the same: they produce image after image and keep the best. Good prompting helps but there is a lot of trial and error.
   
   
2. Find the right service for you. Midjourney users rave about image quality but it’s currently only available on Discord, a social network originally designed for gamers. That can be confusing at first, and there’s a learning curve in prompting Midjourney too: a whole load of special instructions known as parameters. These are very powerful, so don’t be put off trying them but be aware that Midjourney takes time.
   Meanwhile, there are different financial models behind these tools. DALL-E is available only to ChatGPT Plus subscribers, for example, while Image Creator from Microsoft Designer also uses DALL-E but is free to anyone with a Microsoft account.
   
   
3. Know your image jargon. You can use cinematography terms like “wide shot” when prompting an image generator or even tell it to mimic a specific camera lens, such as “taken with a 15mm fisheye lens”. Remember, the AI might sometimes ignore prompt details like those, so be prepared to try a few times.
   
   
4. Balance concision and completeness. As mentioned above, image generators get confused by prompts that are long or dense with detail. Keep your prompts concise and include only the details you absolutely cannot manage without.
   
   
5. Write better prompts with an LLM. One way to improve your prompts is to get another AI to write them. Tell ChatGPT what picture you want to create and then ask it to write a prompt for DALL-E. It might come up with things that you haven’t considered.
   
   
6. Use them to edit existing photos. Microsoft Designer, Adobe Firefly, and similar tools can use AI to edit your pictures. This is handy for removing backgrounds, getting rid of a specific object in a photo or even adding something to the photo.

Staying out of trouble

This is a new technology, so it isn’t without problems. First, it does not understand what it is doing. Just as LLMs are trained on billions of documents without knowing what they mean, so image generators are trained on masses of images without understanding what is in the picture. That’s why they sometimes give people too many fingers or even limbs. They don’t understand objects either, so you might see a picture of a barber ‘cutting’ hair without scissors in their hand.

Words are a particular challenge. Try using an image generator to produce a book cover and you’ll probably get a jumble of repeated letters and vague shapes. This seems like it should be easy to solve, particularly because other types of AI are so good with words, but it’s not. Image generators don’t know they are generating words, any more than they understand hands or scissors. They are approximating details in their training data.

A more serious challenge is accuracy and bias. Google had to apologize after its Gemini AI produced pictures of Nazi soldiers that included black and Asian people. The mistake was easily spotted; most people are aware that these pictures are historically inaccurate and offensive, but not every mistake will be as clear. Businesses must ensure that there is a clear approval process for any AI-generated image they use.

A similar problem comes from intellectual property. You need an approval process in place to ensure that you are not using a service that was trained on copyrighted material. Many artists have complained that image generators are imitating their style. While the legal implications are still being worked out, there is still the potential for reputational damage if someone later claims your business used an image that plagiarized their work. There are potential, deeper ethical issues to consider here too, about AI images depicting real people or that aren’t sufficiently identified as AI-generated.

Related to that is the question of who has rights to your created images. On DALL-E, it belongs to the user, but Midjourney images are free for anyone to use, regardless of who created them.

If that sounds like a minefield, don’t worry. Remember, this technology and the norms for its acceptable use are still developing. Some time playing around with generative AI tools and exploring how they fit your needs is a good way to formulate processes and policies to share with employees. There is enormous potential here for helping your employees save time as well as expanding their creativity.