AI image generators are amazing, but they aren’t very smart. They produce unbelievable results with zero effort. It’s like taking photos with the best camera phones. They also like to show off their inability to count. These are observations I made during my first encounter with generative AI images. I tested 10 of them in search of a favorite, and that was as fun as it sounds.
While looking for the best AI image generator, I learned a lot about them. I noticed peculiarities, ran into issues, stumbled into moral dilemmas, and entered rabbit holes I did not suspect existed. This changed how I see AI image creation, and I’d like to share my impressions with you.
7 AI images don’t have to be perfect to be useful
Especially if you have Photoshop
Can you spot the error in this AI-generated image?
I expected AI images to be immaculate every time, but reality struck. Once in a while, images included obvious imperfections, giving them up as AI-generated. I noticed this could be anything from poorly rendered patterns and textures to people with extra fingers or in unnatural positions.
I realized that flawed images could be useful in some applications. Glitches weren’t visible when the output image was scaled down. Flaws also matter less if the image isn’t in the focus of my content or if it won’t be seen by anyone but me. Here are examples of how I’ve used AI-generated pictures since then, both for fun and at work:
- Make collages and YouTube video thumbnails. In these cases, I downsize the AI-generated picture or use it as part of a larger design.
- Generate slideshow visuals matching the theme of the content.
- Make clip art and backgrounds for videos and animations.
- Visualize ideas like room arrangements, color or material combinations, and concepts for products I can make in my workshop.
When an image doesn’t suit your needs, edit out bits of it or ask the AI for a new one.
6 AI images don’t have to be perfect to cause trouble
Not everyone can identify computer-generated art
A tech enthusiast may be familiar with AI-generated media. They may also know how to recognize images made by AI. But their cousin doesn’t and would not hesitate to share such content with their book club.
“We found that media generated with current state-of-the-art methods has become virtually indistinguishable from ‘real’ media. Across all countries and all media types, people rated AI-generated samples as more likely to be produced by a human than a machine,” states one survey conducted among 3,000 people in 2022.
Surprisingly, even images I can easily identify as AI-generated look convincing and influence unsuspecting people, judging by some of the comments I see online. This is worrying as it creates the potential for misuse. Bad actors can grow social media pages using controversial AI media to rake in likes, comments, and follows. When these pages are big enough, they can be leveraged for shady purposes.
5 Image generators aren’t always smart
I’m not sure if I should laugh or cry at these hallucinations
AI can produce weird images because it doesn’t understand physics, anatomy, etiquette, or what a plumber does. Above are the silliest examples I received during my testing, but the surprises didn’t end there.
I was shocked to realize that AI image generators, as amazing as they are, may fail to process instructions a 3-year-old would understand. Results vary from one generator to another, but it’s common for AI to be unable to process instructions related to object placement or orientation. Imagine drawing one object to the left or right of another.
Counting is also a challenge for AI image generators. The AI may get it wrong when you request more than three things. Images, including text, will likely have extra or missing letters or the words may not make sense. However, there is at least one image generator that renders text reliably. This leads me to my next point.
4 Every AI image generator has its pros and cons
It isn’t only a matter of which is best
I’ve played with more than 10 AI image generators. While Google Gemini is my favorite, I switch to other AI tools for some requests because all have their strengths and limitations.
Gemini, for instance, is free, simple to use, and usually produces great results. However, rendering people requires a paid subscription, and it doesn’t generate images of children. Adobe Firefly has controls that let you set parameters like style or visual intensity, and its ability to use a reference image can be helpful. On the other hand, it struggles with images of people. If you need text in your image, Ideogram outperforms these big-name AIs.
You may need to try several AI image generators to see which best fits your needs. Don’t assume they’re the same because they look identical.
3 Using the optimal text prompt helps
The more specific you are, the better
AI can interpret your text prompts. However, it can’t read your mind (yet), so it doesn’t exactly produce what you’re imagining. Still, you can get results closer to what you have in mind. You must be more specific and provide an optimally structured prompt.
What does an optimal prompt look like? It depends on the AI tool you use. In any case, a short prompt leaves room for improvisation on the AI’s side. However, the AI produces a more accurate image when you provide more details with a good structure. Prompts can get surprisingly long. Trust Insights has a PDF guide that shares best practices and provides example prompts tailored for popular AI image generators.
2 AI image generation uses a ton of power
So, tech giants are going nuclear
Some light implementations of AI run locally and are processed by the NPU in your phone or PC. Image generation is rarely one of them. If you ask Google Gemini for an image, your prompt is processed by a remote server, which requires a lot of processing power.
How much power? A recent estimate by The Verge claims generating a single image can consume as much energy as held in a typical smartphone battery. It isn’t a surprise that top tech giants like Meta, Amazon, and Google are looking to nuclear power to meet the energy demands of their AI projects. This is one reason AI tools are usually paid. Other reasons are the infrastructure and AI model training required to maintain the service.
1 AI image generators are in an ethical and legal gray area
Copyright? What’s that?
Teaching an AI model to generate images requires training it on data. High amounts of quality data yield better AI models. Where does that data come from? Even if that data is publicly available online, it’s debatable whether AI models should be trained on it.
Many of the images we see online are made by people who do this for a living, such as artists and photographers. Training an AI to recreate objects, scenes, compositions, and styles could be unethical. The original creators are not compensated for indirectly helping create a tool that could replace them.
Making matters worse, generative AI can violate intellectual property rights. As of this writing, Gemini would not hesitate to create an image of Sonic the Hedgehog shaking hands with Super Mario.
As the popularity of AI tools grows, so will the number of lawsuits and controversial cases surrounding these gray areas.
AI is powerful, and even more so when you’re good at using it
Some of the best AI-generated images I’ve made with Google Gemini.
It’s easy to overestimate the capabilities of AI image generators when you see a gallery of cherry-picked results. It’s also easy to swing in the other direction and underestimate AI when you run into its flaws and limitations.
Generative AI is here to stay in one form or another. After weeks of using AI image generators for various tasks, I saw them not as complete replacements for me as a creator but as companions. They can lay the foundations for and fill in the details of my creative work or get me out of a creative rut.
AI image generators won’t replace human artists anytime soon. However, humans must adapt to living with AI, which will get better. You can get up to speed on other AI tools with our in-depth Google Gemini and Samsung Galaxy AI guides.
