AI Turns Wi-Fi Signals Into High-Resolution Images With Stable Diffusion 3

With the help of Stable Diffusion 3, researchers at the University of Tokyo have discovered a way to use AI and Wi-Fi signal information to map a real-time image of a room and the people or objects within it. This implementation, called LatentCSI, uses a combination of existing Wi-Fi Channel State Information (CSI) and Stable Diffusion to capture images based solely on Wi-Fi signals, and the visuals are terrifyingly accurate. There are some caveats in play here, especially if you want to get a truly accurate image of a given space, but it does go to show just how much information can be gleamed from invisible Wi-Fi signal bounces with the right tools.

So, what should we make of this? First, let's show you all some of the examples, break down exactly how this works and how Stable Diffusion 3 was trained to produce these images.

The fidelity of the final picture here is an approximation, but also just goes to show how far existing technologies like Wi-Fi CSI had already brought us toward next-generation surveillance that doesn't even require a camera. After all, Wi-Fi CSI has been used for motion sensing capabilities as far back as 2019, with the Linksys Velop Tri-Band AC2200 routers.

But how on Earth could a Wi-Fi signal by itself show you what the inside of a room and what its occupants look like?

In short, it doesn't. Stable Diffusion 3 was not only trained on Wi-Fi CSI in order to produce these images, but also photographs of the rooms themselves and potentially even the occupant upon which LatentCSI was being tested. There is still impressive tech at work here, to be sure—translating Wi-Fi CSI pixel space imagery into AI-compatible latent space imagery—but if these images require existing photographs of the spaces being monitored, its applications are limited. Wi-Fi CSI could already track how objects were moving in a given room before this advancement was made—all Stable Diffusion 3 is doing here is turning that reading into something more visually cohesive than a noise map formed by bouncing radio waves.

So, is this the end of privacy as we know it? Not any more so than Wi-Fi motion sensing capabilities already were by themselves. However, the ability to train models in this way does show a potential advancement for existing surveillance solutions, where it could be used in areas without an active camera or that are otherwise just outside of a camera's range. That's more likely to be limited to private business and government facilities, though. Still, it's kind of spooky to see these kinds of advancements in AI-enhanced surveillance technology.

Image Credit: Eshan Ramesh, Takayuki Nishio, School of Engineering, Institute of Science Tokyo via Arvix.org