``Mist'', a tool for adding electronic watermarks that interfere with the learning of image generation AI for your own illustrations and photos



Image generation AI such as Stable Diffusion and Midjourney learn from the images in the dataset, but there is a debate about the copyright of the images included in this dataset, and people say, 'I don't want my illustrations and photos to be learned by AI.' There are many people who say. ' Mist ' announced by a research team at Shanghai Jiao Tong University is a processing tool that prevents AI from recognizing the image by adding an invisible watermark to the image to prevent learning.

Mist

https://mist-project.github.io/index_en.html

[2302.04578] Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples
https://doi.org/10.48550/arXiv.2302.04578

GitHub - mist-project/mist
https://github.com/mist-project/mist

The research team has published an image on the website that shows how much learning can be hindered by actually using Mist.

The following is the case where Van Gogh's 'Sunflower' is read with 'Textual Inversion ' that can output results close to pre-loaded images. 'Source image' is the original image, 'Image Generation from source image' is the output result image, 'Misted image' is the original image watermarked with Mist, and 'Image generation from Misted image' is the misted image. This is an image of the result of reading and outputting. At first glance, there is no difference between the Source image and the Misted image, but the result of loading the Misted image looks like a mosaic pattern, and it looks like the image is completely broken.



In the case of '

Dreambooth ' which additionally learns a small number of images, it looks like this. Normally, Van Gogh's brush strokes are reproduced, but when Misted image is read, a vague image with noise is output.



When you load the image into the image generation AI ``

Scenario.gg '' that creates assets for the game, it looks like this. When Misted image is loaded, an image is generated that completely ignores the pattern and composition.



When the misted image is loaded into '

NovelAI ', which can generate anime images with high precision, the overall color scheme is inherited, but there is still a large amount of noise, and the picture is not practical.



When Mist is used with the default settings of 100 steps and 512 x 512 pixels, it is said that images can be processed within 3 minutes, and high-speed processing is appealing.



Mist can be used alone, and can also be used with the Stable Diffusion web UI . Mist is licensed under the GPL-3.0 license, and Windows and Linux versions are available on HuggingFace.

mist-project/mist-win at main
https://huggingface.co/mist-project/mist-win/tree/main

mist-project/mist-linux at main
https://huggingface.co/mist-project/mist-linux/tree/main

Related Posts:

in Software, Posted by log1i_yk