The new version of Illuminarty is now ready to launch!
First of all, thank you so much for the support, starting from when we made our prototype last year. Your support was the sole reason why we could keep on working on this project.
We would like to share here some of the changes we have made to the prototype of Illuminarty, some of the known problems, and how we are planning to tackle these problems.
The previous version of Illuminarty was a single-page web application that only had an image detector, and served a simple API that could only handle a few requests. However, to scale up the service and improve the detection performance, the operating costs were expected to go over our budgets very quickly and thus make the project unsustainable. Our goal was to provide the AI detection service for free; we could not push further only to stop maintaining our service.
To this end, we decided to add new features that are provided on a monthly subscription. As you would have already seen from our current features page, there are 3 additional features for image detection that you can access by subscribing to our Basic or Pro plan:
- Model Attribution
- Image Localized Detection
- API Access
We would like to note again that the basic detection functionalities will always remain free. These additional features, or equivalently the monthly subscription plans, are made to support the growing server fees and training costs for detection. We will improve the infrastructure and the performance of the detectors accordingly as we attain revenue through monthly subscriptions or donations to our Patreon.
upgrades on image detector
We also made some updates to improve the performance of our image detector. Until now we were only able to selectively sample the images from a few well-known generation models, but after the launch, we plan to set up a proper pipeline for the continuous training of a more diverse set of generators and to include different art styles and media of human artists.
We share below the comparison of the performance (validation accuracy) of our detectors within our collected dataset:
|Detection Task||Previous Model Accuracy (%)||New Model Accuracy (%)||Difference (%p)|
|AI Detection (binary)||91.8||97.5||+5.7|
|Model Detection (multi-class)||90.1||96.4||+6.3|
|Stable Diffusion (and its variants)||79.1||93.5||+14.4|
However, as we will discuss in more detail in the Known Problems, these numbers need not match the actual experience of the users. This is because our current collected dataset does not cover the diverse styles and tools of the human artists, the detector currently tends to pick up some of the drawing tools, e.g. automated brushes, as artifacts coming from an AI generator.
new text detector
We have implemented a proof-of-concept for text detection on the new website. Since it is yet a proof-of-concept, there is a lot of room for improvement and thus the results should be taken with caution.
Nevertheless, the reason why we have released the text detector on the website is to show that we are willing to fight against the problems occurring in various media of AI generation, not only for the images. We believe that in the future we will need to use every piece of information that is publicly available to discern between the behaviors of AI and humans. Thus the implementation of the text detector was our portrayal of such a statement.
We share here some of the known problems for the current version of Illuminarty, and our plan on solving the problems.
Cannot process large images (> 3MB)
- We will lift the size limit of the images as we improve our infrastructure.
Fails on specific art styles, e.g. color schemes, line styles
- We assume this is because of the bias in our sampled dataset. We believe that gathering more data will naturally solve this problem.
- But to collect the right representative data points your feedback is crucial. Please let us know if you are getting wrong outputs, we will note the modes of failure and remove them in the next updates.
Fails on high resolution images
- This seems to come from various factors. The most dominant factor seems to be that most of them are artworks that use automated brushes and automated light expressions, and our image detector seems to pick them up as an artifact coming from an AI generator.
- Again this requires us to collect sufficient amount of samples that use such tools. Please let us know if your works are deemed as AI generated, it helps us a lot to understand the errors.
Fails on photographs, anime and game screenshots
- The above media tend to undergo a lot of post-processing, most of which use AI algorithms for rendering and upsampling. Thus in most cases they leave similar traces as the AI generators.
- However, we acknowledge that there is a semantic difference between these processing techniques and the AI generation, and thus should be distinguished. We would like to discuss with you where the boundary should be. Please reach out to us if you are getting errors in these domains.
If you are experiencing an issue that is not on the list above, please let us know through our Discord or any of our other social media channels.
We would like to conclude this post by sharing the new features we are preparing for the next update.
We used to serve the website in 4 different languages (Chinese (Simplified), English, Japanese, Korean) in the previous version of Illuminarty. We will proceed to add them again on the new website, both in the homepage and the web application.
Please let us know if there is another language that you want us to add on the website, and if you could help us out on the translations.
We are working on a web browser extension which let us use our detectors while we surf on the internet.
The extension will communicate with our API and return the AI generation probability for both image and text.
Stay tuned for our latest updates, and tell us your ideas through our social media channels below.