AI in Visual Content: A Guide to Safe and Smart Image Generation

Today, it is hard to miss the impact of generative AI, whether at work or in our personal lives, reading the news or browsing social media. And one particularly prominent area is AI image generation.

You may have checked out Bayer’s customized generative AI tool that works specifically with Media Pool images. Or, you have already been using public AI services such as Midjourney, Stable Diffusion, Dall-E or Adobe Firefly. Either way, interest in AI-generated imagery is rapidly growing.

That’s why we have produced this brief introduction to AI image generation, its benefits, and its risks. If you are new to AI, this article should give you the background to help you get started confidently. If you are already experienced, there are tips throughout and at the end to help you sharpen your skills. 

How does AI image generation work?

AI tools can generate images – either from scratch, or, as in the case of the Media Pool GenAI tool, by building on an existing image – based on text prompts you give them. These prompts could be very simple (e.g. ‘egg’, ‘flower’) or can be more complex and include different factors (e.g. ‘cartoon image of an older man with a beard standing on a pier, evening’).

It’s important to understand that the AI model does not ‘draw’ images like a human would. Instead, it has analyzed millions of existing images and their text descriptions (known as its ‘training data’), and from this, has identified common patterns of pixels, colors and shapes. So, just like an AI language model does not really think for itself, an image AI does not technically know what an egg or a flower is – but it can produce combinations of pixels that match what is found in images of eggs and flowers online, thereby creating something that looks believable to us.

This may make AI image generation sound rather imprecise. But when you start trying different prompts in any major image generator, you quickly realize that it can be very impressive. Generative AI is a huge focus for the tech sector, and models are always being refined and updated with new data. That means AI image generators can faithfully replicate a huge number of image types, compositions and styles – and they continue to change and improve.

Check out the examples below for an idea of what AI can do (and some pitfalls to watch out for)… 

Some (good and bad) examples of AI-generated imagery

Tool used: Midjourney

Bad & mediocre examples:

Good/adequate examples:

Using AI imagery: basic guidance and compliance 

So, can (and should) you use AI image generation at work? The short answer is: yes, with some conditions.

At Bayer, we are committed to continually evolving, up-skilling ourselves and harnessing the latest technology. And we see unlocking the possibilities of AI as part of that. Therefore, it’s permitted for colleagues to use AI image generators as part of their professional role.

Also, outside of a number of specific possibilities in the healthcare sector (e.g. AI systems that relate to medical equipment or biometric data), most of our generative AI use is expected to be ‘general purpose’ and therefore not covered by the responsibilities that relate to ‘high-risk systems’ in the EU AI Act.

That said, we do not have to use AI in every circumstance – and our use should adhere to Bayer’s GenAI Guidance and Rules. That includes respecting copyright and data privacy, avoiding stereotypical or demeaning depictions of people or themes, and being transparent about our use of AI (labelling images accordingly, and certainly never pretending we did not use AI when we did, or using AI in a customer project when it’s not specifically approved). Check the links in the sidebar for the full guidance and rules.

Getting the most out of AI imagery 

Good AI practice isn’t just about compliance: it is also about maximizing utility and quality. It’s easy to get some negative ideas about AI image generation: that it is only useful for a certain type of image (e.g. uncanny almost-photorealistic imagery), or is only for adding to social media posts, or is a ‘quick and low-effort’ route when you have something in mind but can’t find a better option. And the amount of generic AI art that you find online probably does not help. But when we go beyond these misconceptions, we can find new use cases and possibilities that can transform what we get from AI image generators.

For example, AI images do not have to be of ‘real-life’ subjects. Try prompting to produce templates for diagrams or infographics, so you can use them as a starting point in presentations, or to brief creatives. It can help you visualize your ideas and save time. 

Also, the creative process does not have to end with an AI prompt! Some of the best ‘AI-generated’ images will have a round of human editing and retouching afterwards. Or, even, the AI image was used as inspiration for something created separately. 

Once you get the hang of it, you can use AI image generation for a lot more than one-dimensional prompting: ideas, examples, communicating concepts, combining with other AI tools, or going beyond your ‘comfort zone’ and showing you things you would never have come up with yourself.

The downsides of AI image generation

At the same time as exploring the enormous possibilities of generative AI, we should also acknowledge the criticisms of it. Issues with AI that often come up in public discourse include:

• Copyright and intellectual property issues: AI models are often trained on copyrighted images, leading to some high-profile legal cases being brought by artists against AI providers.

• Sustainability: The energy and water consumption associated with prompting an AI engine is understood to be much higher than with a regular online search.

• Bias: AI models reflect the data they are trained on. We cannot assume this data is unbiased, and should watch out for AI-generated images that perpetuate stereotypes or lack diversity in the people they depict.

• Privacy Violations: AI can generate identifiable images of real people. If they haven’t consented to this, it’s a privacy violation and a potential legal issue (commercial use of someone’s likeness without a release). 

• Effect on creative jobs: There are fears that AI-generated imagery could be used as a complete or partial replacement for human photographers, illustrators, designers and other creative professionals – affecting their job security and removing the ‘spark’ of human creativity from workflows.

• Truth and authenticity: Since AI-generated images can sometimes be indistinguishable from real ones, it creates a risk of misinformation being spread, or of undermining people’s trust in otherwise reliable sources.

Before proceeding with external uses of AI-generated images, it’s recommended to consider how these downsides will be avoided or mitigated. Consider the points above and the tips below, and stay realistic in your assessment of AI image generators. We can be excited about them without believing all the hype! They are not infallible, and are only as good as the data that is put into them (either in training or through your prompt). And they do not, and should not, devalue the work of human creatives, which is as important as ever.

At its best, AI image generation can be creative, fun and inspiring – so let’s get involved and use it in the right way. And if you have used AI images professionally in a way that you are proud of, please let us know so we can share the example!