I’m CNET’s AI image and video generator reviewer, and one of the best parts of my job is laughing at the truly terrible, deeply flawed, occasionally frightening AI images that pop out while I review different AI image generators.
Don’t get me wrong, AI creative software has come a long way in a short amount of time. I’ve created beautiful sci-fi scenes in Midjourney, realistic looking product images with Stable Diffusion and AI-ified myself with ChatGPT. But no AI service is perfect, and I’ve noticed some trends throughout my testing. There are some things that, despite constant updates, AI image generators simply can’t handle well on the first try.
Some AI image programs give you tools to edit these mistakes, which is great. I highly recommend to take advantage of them when you can. But for programs that don’t offer them, or more likely, when they don’t fix the problem, this is what you should do next.
Based on my experience, these are my tried-and-true tips for tweaking your prompts and settings to fix your images. For more, check out the best AI image generators and our guide to effective AI image prompt writing.
10 Photos That Show What AI Image Generators Struggle With Most
See all photos
Human faces and expressions
Accurate facial expressions continually challenge AI generators. Quirky eyes, teeth and eyebrows are some of the strongest indicators that an image is AI-generated. In this case, the result was extremely funny to me, if also completely unusable. The girls are sporting some Halloween-like vampire teeth, and the dude in the back is having more than a bad hair day.
Even with cartoon or non-realistic characters, generators struggle to moderate emotion and expressions. This image — created by our best pick, Dall-E 3 — over-amplified the prompt, and the end result was too dramatic. I’m a self-identified neat freak, but I can’t imagine anyone getting this upset over what looks like hundreds of dollars of cleaning supplies. Even the best programs can fail and produce wonky results.
How to fix it: I recommend asking the service to cut down on the number of people it’s trying to render — cut down on the number of chances for error — and using post-generation editing tools to select specific parts of the image that need regenerated or fixed. Picking a more mild adjective (“angry” rather than “enraged”) might help guide the service down the right path.
Iconic characters, trademarks and recognizable logos
For many reasons, AI image generators can rarely produce accurate logos, known trademarks or otherwise recognizable characters. This is usually because there are significant legal concerns about a company allowing its brand materials or intellectual property to be used for AI services. It also could be because the logo or brand you want to include isn’t in the generator’s training data, so the generator won’t understand your request.
Two recent exceptions to this are new image generation features on the Google Pixel 9 and Pixel 9 XL, thanks to Gemini AI, which our mobile experts used to create images featuring accurate-looking Mickey Mouse and Pikachu renditions. The other exception is some paying X (formerly Twitter) users reporting that they can use the social media site’s Grok chatbot to make realistic-looking Mickeys and human political figures. These images still aren’t perfect renderings, but they’re pretty close.
How to fix it: You can’t, but on the plus side, you’re a lot more likely to avoid lawsuits for copyright infringement. If you run into this issue, I recommend rethinking your design concept to see if you can get the same point across without using a specific company name or logo. Do you really need the TikTok logo, or do you just need a phone with a vertical video playing, for example.
Overlapping, complex elements
Occasionally, you’ll stump a generator if you have too many overlapping or complex elements within an image. The best AI image programs aren’t going to run into this issue very often, but even the best programs fail at some point. With this Leonardo image, the library of my dreams is ruined by the rolling ladder that disappears halfway up.
Similarly, this photorealistic image of a kitchen looks alright from a cursory look, but a second glance or quick zoom-in shows nonsensical characters printed on the cookbook, and the book itself seems to have two spines and three sections. Images like this show how frustrating it can be to create AI images — even when you think you’ve got a winner, small flaws make them unusable.
How to fix it: Try simplifying your prompt or using post-generation editing tools to single out the problem and fix it. Sometimes changing the aesthetic can help as well; these issues usually pop up when you’re trying to make photorealistic or stock images. If your service offers a specific area editing tool, try using that to select the questionable area and ask the generator to remove what’s in it.
Over-editing and hallucinations
The sign of a great AI image generator is its ability to produce an image with few errors or hallucinations while giving you the right tools to fix those issues when they inevitably pop up, but not every service does. In this example from Midjourney, I went through several rounds of edits on this image of a soccer team celebrating a victory, and I have no idea what happened to get to this final result — or any clue what’s going on with the blob that’s presumably a player in the right-hand side of the picture. Frankly, I don’t think Midjourney knew either.
How to fix it: With editing AI images, sometimes less is more. Don’t be afraid to scrap your current batch of images and start over. You can often preemptively fix big issues by refining your prompt so you only have to worry about fixing minor issues later.
AI generators still need human work
I have no doubt that the companies behind these AI generators are working to eliminate these issues and level up their models. For now, imperfect images show us once again that while this tech is improving, it’s not flawless. My hope is that because so many of these services face similar challenges, they will all be chasing solutions for these issues.
As a final note, I’ll remind you that it’s a best practice to credit or acknowledge that an image is AI-generated when you share it. As AI images get better and more realistic-looking, it’s more important than ever to distinguish AI-generated art from other kinds of media.
For more, check out our advice for writing the best AI image prompt and our best picks for AI image generators.
Read the full article here