Sep 15, 2022 14:00:00

'Have I Been Trained?' allows you to search whether your work has been used without permission by image generation AI

Stable Diffusion , an AI that generates images based on input text, is attracting a lot of attention because it can easily generate images of your choice. On the other hand, some people have complained that the use of ``illustrations drawn by real artists'' during AI learning is a violation of artists' rights , and ``I don't want my work to be used for learning.'' I'm sure there are many people who are worried. The large-scale dataset `` LAION-5B '' containing over 5.8 billion images is used for Stable Diffusion's training, and the web application ` `Have I Been'' allows you to search for images included in LAION-5B. Trained? '' was released, so I tried it out.

Have I Been Trained?
https://haveibeentrained.com/

When you visit 'Have I Been Trained?' you will see a search bar in the center of the page. Search can be performed using text strings and images.

◆Search by string
To search by string, enter the string you want to search in the search bar and press the Enter key. This time, I searched for 'Naruto Uzumaki', which is the Romanization for 'Naruto Uzumaki'.

The search results for 'Naruto Uzumaki' are below. A lot of fan art was displayed.

The search results for Studio Ghibli's English notation 'Studio Ghibli' are like this. You can see that many images from Ghibli works are also learned.

You can search not only by artist name but also by real person's name. When I searched for the name of AMD's CEO Lisa Su, a large number of ``photos of CEO Lisa Su holding the CPU'' were displayed.

◆Search by image
In 'Have I Been Trained?', you can check whether or not this image is being used for training by searching using the image. This time, we will investigate whether 'Mona Lisa' is used for learning.

First, click the camera icon in the search bar.

Next, select the image you want to search.

The search results for 'Mona Lisa' are below. A message was displayed stating that ``An image was found that matched 99% of the searched image.''

◆About LAION-5B
The dataset 'LAION-5B' used for Stable Diffusion training contains over 5.8 billion images and is released for free under the 'Creative Commons BY 4.0' license with the assumption that it will be used for research purposes. Masu. The images included in 'LAION-5B' are randomly extracted from the Internet and are not screened by humans. For this reason, 'LAION-5B' includes many works by artists.

Detailed information on 'LAION-5B' can be found on the following page.

LAION-5B: A NEW ERA OF OPEN LARGE-SCALE MULTI-MODAL DATASETS | LAION
https://laion.ai/blog/laion-5b/