Freepik’s new AI model, Mystic, has just hit the scene. For the past week, it’s been creating a stir among digital artists across the web.
So far, Mystic has made a name for itself with its incredibly lifelike, high-quality images. This is raising concerns for those of us who want to easily spot AI-generated pictures.
The usual AI slip-ups, like extra fingers, are gone or minimal. As a result, it’s becoming harder to identify fakes and catch scammers.
But are these images too realistic to be detected or can the best AI image detectors still thrive?
Let’s find out.
Freepik, a Spanish-founded website for designers to find high-quality visuals, launched its first AI image generator “Pikaso” earlier this year.
Their latest model, Mystic, combines the FLUX.1 model suite with Magnific, an advanced AI upscaler and enhancer, which gives premium users the capability to create high-quality, highly photorealistic images with ease.
One of Mystic’s best features is that it can make very clear, detailed images right on the platform. These images are incredibly sharp, with a 2K resolution.
Early user feedback on X suggests that Mystic currently beats Midjourney in terms of photo-realism and DALL-E 3 in terms of precise prompt representation.
Recently, we tested the best AI image detectors on images generated with FLUX.1.
In that article, we discovered that the tools could detect FLUX.1 images, but only those which were created with the more premium models – dev and pro, but not those generated by schnell.
So how will the detectors perform with images generated by Mystic?
This really sparked our interest as Freepik is a widely accessible tool that will be used by a large population.
For testing, we chose four tools that performed well with FLUX.1:
After that, we generated three images of varying difficulty for the detectors to beat:
Image 1 – Hard-to-detect: A picture of the Scottish Highlands
Image 2 – Moderately difficult to detect: Chess players in a park
Image 3 – Easy to detect: A crowded theme park
So how did it go?
| AI Image Detector | Image 1 – Scotland | Image 2 – Chess | Image 3 – Theme Park |
|---|---|---|---|
| AI or Not | 100% AI | 0% AI | 100.0% AI |
| Is it AI? | 10.0% AI | 80.0% AI | 81.0% AI |
| Hive Moderation | 98.4% AI | 99.9% AI | 99.9% AI |
| Winston AI | 96.0% AI | 10.0% AI | 100.0% AI |
The results were interesting, to say the least.
Hive Moderation absolutely aced this round. It got everything right and with high accuracy. They have also included the FLUX.1 indicator in their analytics, so they are keeping up with these recent AI developments.
Regarding the other tools – all of them made a big mistake regarding image 1 or image 2.
Is it AI? didn’t recognize that the Scottish highlands image was AI-generated – so if you are using this tool, for the moment, try to use it on images where there are more artifacts present. Overall, it performed well.
AI or Not got the chess players wrong – at the moment we don’t have a good explanation for this, but you can bet that the team at AI or Not is working hard to solve this oversight. Otherwise, the tool performed very well.
The same can be said for Winston AI; however, their tool is in beta, so it is expected to improve very soon.
We also evaluated the widely-used Fake Image Detector, which provides text-based results instead of confidence scores. This tool performed worse than the other detectors mentioned above, as it incorrectly classified the first and last images.
Freepik Mystic is a capable model, able to generate ultrarealistic images with ease.
We were interested in determining if we could detect images generated by this model, as we expect it will be widely popular among the general population, even outside the AI artist community.
Our testing showed that Hive Moderation works best to detect Freepik Mystic images. However, the other tools also performed well for certain use cases.
Before AI detectors can catch up, it would be smart to study the results presented in this article, look at our other article about FLUX and Ideogram, and choose the detector that best suits your use case.
It’s also a good idea to use multiple tools if you have suspicions. If two tools give you completely different answers, it should raise some red flags for you.