We saw insight into opportunity in that there was no service that offered what we have developed. Asking around and conducting surveys we discovered interest in the idea, some artists suggested adding features, one of them being a masking feature to hinder AI from copying media on the site. While masking technology does exist, most of the artists we spoke with did not use it personally. This insight revealed a gap in the market, allowing us to automate the masking process and reach an untapped audience.
The idea is to develop an app/platform that stores a piece of media, and then allows for (people with access) to edit said media somehow, and add their attribution in addition to the original maker's in the updated piece of media. And we preserve these file versions and attributions for every update anyone makes. And on top of that we try to make it difficult for AI to steal from these media pieces when generating similar content, though that's a side feature and not the main feature
We collaborated on Anvil using Visual Studio Code, Live Code, Github.
The main challenge involved the masking system generally, from which to use, if it could be automated, if it could work in a reasonable amount of time, and if its masking actually hindered AIs. We went through multiple different candidates, ranging from applications, to open-source research tools. All had different strengths and weaknesses but all most all would have been extremely hard to run with a script or took too long to process images. The open source masking tool Mist was by far the strongest pick, being able to run in the terminal, with a fairly short run time of 3 minutes a image with a strong masking effect. The taxing part of adapting Mist was its subpar documentation, and errors with python requirements. Despite the challenges we created a script that flawlessly runs Mist. We had planned on training on our model, but ran out of time.
Automating Mist to take a given media from our database, run the masking process, then uploading the image back online.
What we learned. Used Nightshade as a foundation in understanding ways to combat AI theft.
What's next for Anvil. Refining the automated process of masking media to be faster and more powerful as the base library for masking develops further and further.
Log in or sign up for Devpost to join the conversation.