AI Tagging of NSFW Images and Videos

In 2019, I created the model behind Puryfi, a state-of-the-art AI that detects and locates nudity and body parts in images. Ever since completing that model, I've envisioned the potential of integrating AI with NSFW images and videos. My goal has been to develop one or more models that, when run together on an image or video, provide enough data for a person (or computer) to understand the key elements without viewing the content itself. I refer to this concept as "Content Understanding".

Perfect Content Understanding of an image or video would enable many potential use-cases:

• Enhanced Search Functionality: Improved search capabilities by leveraging detailed and accurate tags.

• Automatic Tagging: Save time by automating the tagging process for large media libraries.

• Time-based video tagging: Users could see exactly what's happening at any point in time in a video and skip to specific parts.

• Metadata Enrichment: Add detailed metadata to your media files to improve the overall organization.

• Content Recommendation: By understanding the content of images and videos, we can recommend other content you might like. The better our "Understanding" of content, the better we can predict your preferences.

I've attempted this project several times since 2019, but it always seemed unachievable until I made some key breakthroughs. These breakthroughs have led to the current model and proof that this concept is achievable.

Perfect Content Understanding is a complex project that will require years of development and multiple AI models. For now, I've managed to create the first AI model in what will be a set of many, each contributing to the goal of content understanding.

The first model being released focuses on tagging "Sexual Actions" in images and videos. While this model is not perfect, it is getting pretty accurate. See the demo for some examples.

In addition to the model itself, I’ve developed a plugin for Stash, a popular open-source media organizer tool. This plugin integrates seamlessly with Stash to provide automatic tagging and enhanced content understanding directly within your media library.

As mentioned earlier, achieving truly perfect Content Understanding will take years of development. However, I believe this initial model will be useful to many, and I want the community to help drive the priorities for future developments.

Please be aware of the current limitations before becoming a patron.

Here is the full list of tags that this model can predict:

https://docs.google.com/spreadsheets/d/1MbULbXaUMjvuGo066SuPaJVjBSSA57jhS3DAtvWNpwI/edit?usp=sharing

I've been excited about the possibilities of this project for a long time, and I'm thrilled to hear how these models help users with their content. I'll be eagerly awaiting your feedback as you start using it and discovering new ways to leverage this technology!