We clarify the key effect of Mist-V2 and its constraints in the following statement to avoid misunderstanding:

“Apply this noise. Mist-v2 may raise the cost of a motivated and professional attacker. And it might help against someone who’s not really trying or not an expert. Mist-v2 is not a perfect solution, but it would be certainly better than nothing.”

What is Mist

Mist is a powerful image preprocessing tool designed for the purpose of protecting the style and content of images from being mimicked by common AI-for-Art applications, including LoRA, SDEdit and DreamBooth functions in Stable diffusion, Scenario.gg, etc. By adding watermarks to the images, Mist renders them unrecognizable and inimitable for the models employed by AI-for-Art applications. Attempts by AI-for-Art applications to mimic these Misted images will be ineffective, and the output image of such mimicry will be scrambled and unusable as artwork.

Mist

Mimic

We are committed to developing and maintaining Mist in a long term and continuously enhancing its function, and to this end, we have open-sourced Mist on GitHub. We hope to foster a vibrant community for developers and users on Discord to collaboratively improve the performance of Mist. We welcome both user responses and technical contributions. Join our community and this exciting endeavor now!


News & Updates:
[2023.12] Mist V2 is officially released as the first image preprocessing tool that has been systematically verified to be effective under LoRA. Furthermore, it significantly enhances its capabilities against various AI-for-Art applications and reduces the computational resource consumption.
[2023.04] Mist is also accepted by ICML 2023 as Oral Presentation! See our paper for more details.


Advantages & Examples

Mist is effective against a variety of AI-for-Art applications and is highly robust to noise purification, and takes less time and less computational resource to function. The following examples show that Mist provides the most advanced protective watermarking for images that can effectively resist various noise purification methods such as cropping, resizing and super resolution. In addition, the process of adding a watermark to an image by Mist takes only a few minutes.

Effectiveness close

Mist is effective against various AI-for-Art applications, including LoRA (implemented in Diffusers), SDEdit (implemented in stable-diffusion-webui), DreamBooth (implemented in Diffusers), Scenario.gg.
Taking the two most typical AI-for-Art applications LoRA and SDEdit as examples, the watermarks added by Mist interfere greatly with the results generated.

LoRA

Source image

Image generation from
source image

Misted image

Image generation from
Misted image

SDEdit

Source image

Image generation from
source image

Misted image

Image generation from
Misted image

Robustness under image transformation close

Mist is robust to image transformation. We compare Mist's performance under various defenses such as Gaussian noise, JPEG compression, resizing and super-solution. It is observed that Mist maintains a high level of performance after input transformation.

LoRA

Image generation from Misted image under Gaussian noise

Image generation from Misted image under JPEG compression

Image generation from Misted image under resizing

Image generation from Misted image under super resolution

SDEdit

Image generation from Misted image under Gaussian noise

Image generation from Misted image under JPEG compression

Image generation from Misted image under resizing

Image generation from Misted image under super resolution

Time efficiency close

Mist supports both CPU and GPU. Adding a Mist watermark to a single image using CPU takes approximately one hour. Using GPU, Mist can run with only 6GB of VRAM and complete the processing of an image in just 5 minutes on average.

User cases close

The following are showcases of the results when users employ Mist to protect their artworks.

@桑德兰的等待

Source image

Image generation from
source image

Misted image

Image generation from
Misted image

Source image

Image generation from
source image

Misted image

Image generation from
Misted image

@Anonymous Artist

Source image

Image generation from
source image

Misted image

Image generation from
Misted image

Source image

Image generation from
source image

Misted image

Image generation from
Misted image

Download

Mist is compatible with both Linux and Windows operating systems. For local deployment, users can refer to our guidance for quick start. Windows users can download the Mist launcher from the provided Google Drive or Baidu Netdisk link (Extraction Code: m4nx) and install it for use. For Linux users, we recommend obtaining Mist from our open-source code on GitHub and referring to the Readme file for installation and usage instructions. Additionally, a Colab Notebook is available for users with MacOS systems or those who do not possess proper Nvidia GPUs.

Our vision

There is no doubt that the AIGC applications will revolutionize the production and consumption patterns of all mankind. However, its impact on the established social pattern and the distribution of interests should not be overlooked. We recognize that technological innovation is just as important as the improvement of social systems. While it often takes time for new systems to develop, the attention and reflection should be forward-looking.

Mist and our future projects aim to engage in a form of social practice in a technical way, exploring the possibilities of integrating relevant technologies into the society in a more sustainable and gentle manner. In the case of Mist, we hope to draw attention to the challenges that AI-for-Art apps pose to the established copyright system, their impact on the commercial and aesthetic value of the artist community, and in essence, how we perceive and incentivize human creativity. We’re also actively working on additional projects aimed at addressing ethical, copyright, and trustworthiness concerns arising from AIGC technology.

We warmly welcome all developers, researchers and practitioners who are interested in this vision to contact us, and look forward to exploring the current technological bottlenecks and unresolved potential societal issues together. Reach out if you have:
- Ethical concerns about AIGC to be resolved.
- Technology solutions or ideas around trustworthy AIGC.
- Thoughts beyond tech – your observation on industry practice and AIGC’s social impact.

FAQ close

Q: Compared to Mist V1 and other similar products, what are the main advantages of Mist V2?
A: In the months following the release of Mist V1, we identified two main issues in practice. First, LoRA gradually became the primary method for AI to learn artistic styles, but existing watermark tools were unable to address this problem and lacked the capability to protect against LoRA. Second, existing watermark tools required significant computational resources, often beyond the capacity of artists’ computers. In comparison to existing watermark tools, Mist V2 has made significant advancements in addressing these two issues. Mist V2 is the first watermark tool systematically verified to have protective functions against LoRA. Besides, the computational resource consumption required by Mist is also significantly optimized, allowing it to run on GPUs and CPUs with a minimum of 6GB of VRAM.

Q: How does Mist V2 protect against AI-for-Art applications like LoRA and SDEdit?
A: Mist V2 mimics the training process of AI-for-Art applications such as LoRA and SDEdit on artistic works. It strategically introduces misleading noise into the watermark to confuse AIGC models between the actual content of the artwork and the chaotic patterns embedded in the watermark. At higher intensities, the model may perceive the chaotic patterns as part of the artistic style, resulting in images with the chaotic patterns. At lower intensities, the model struggles to accurately learn the artistic style, reducing the similarity between the output images and the artworks used for training and the diversity of the output images.

Q: Will Mist V2 be bypassed?
A: Currently, there are no simple, lossless bypassing tools available. Any tool that significant undermines the protective effect of Mist V2 would also compromise the quality of images, thereby equally affecting the output quality of artistic style imitation. Nevertheless, considering the rapid pace of technological advancements and iterations, we cannot predict whether new artistic style imitation techniques and models might bypass the protection of Mist V2. However, similar to this update, we are committed to continuously updating and maintaining Mist, addressing newly emerging issues over the long term.

Q: If I use an attack with intensity lower than the recommended level to make the watermark less noticeable, will Mist still be effective?
A: The recommended intensity of Mist is designed to ensure comprehensive performance in various imitation scenarios. If a user decides to decrease the watermark intensity, Mist can still provide protection through a relatively diminished performance. Additionally, Mist allows users to choose different protection modes for various imitation situations, allowing them to adjust the watermark intensity to better suit their needs. Users also have the flexibility to customize Mist’s intensity. For a detailed guide on selecting the appropriate mode, please refer to our documentation.

Q: How does Mist V2 protect against AI-for-Art applications like LoRA and SDEdit?
A: The Mist series watermark tools will permanently be free and open-source . As technical developers, one of the most impactful ways to support us is by starring our projects on GitHub (it's really important to us!). Likewise, we warmly welcome users and developers to provide feedback and report any issues with Mist through our Discord community or other contact methods, helping us further refine the technology.
We also greatly appreciate your expression of interest, praise, and support through social media. In addition to releasing more AIGC-related technical projects, we plan to introduce and inform the public about the technical principles, business practices, and potential societal issues related to AIGC through social media in the future. We hope to promote a more comprehensive and in-depth understanding of AIGC and related technologies through sharing. Follow us to ride the wave of technological change together!

Contact us

Mist is led by Psyker Group, a group of developers and industry practitioners committed to work on technology projects addressing trustworthy, ethical and regulatory issues arising from AIGC tech nology.
Team member of Psyker Group:
Caradryan Liang, Nicholas Wu, Chris Xue, Melo Yang

Main developers of Mist:
Psyker Group, Boyang Zheng, MOSS星辉

We extend our special thanks to Alice2O3 for the exceptional contribution in developing the launcher for Mist.

Yongwen Su at UCB also contributes to the development of Mist. We also thanks Jiahao Wu, Yi (David) Zhao, Jiahao Wu and Yi Zhao for their advice in software development. We would also like to express our gratitude to 苹果, GUUUU, 原野, 蚕蛹子 and BASS for their support of our project.

We value user and developer feedback. Join our Discord community to share thoughts, ideas, and suggestions for Mist's improvement. Our team is available on Discord to address Mist-related issues.
Updates on Mist will be posted on @Psyker_ (Weibo) and @psyker_202304 (Twitter).

You may also contact us via QQ group chat: 189980587 or email: mist202304@gmail.com.