High-Quality Video Reconstruction from Brain Activity

Not gonna lie, it's impressive. But that said, courtroom drama this is classic recovered-memory stuff:

"your recovered memory says you saw TWO people walking. Police footage shows ONE person walking"

"case closed: its a false memory"

So its prompted adversarial semantic match is good: birb matches birb. jogger(s) match jogger. cloud matches cloud. But its not faithful image recovery at any stretch. (nor do they claim it. It's how the downstream consumption of this work will what-if on it)

Some of the goodness looks like pure and simple motion recovery: if you put cat == cat and then it's picked up tracking, making it tracking-cat isn't exactly hard.

This sounds and looks impressive but is this actually anything more than a brain machine interface for a stable diffusion like model using an MRI?

If so what it generates and reconstructs is based on the data the model was trained upon not what the individual has actually seen. So this only provides a thematic prompt at best and the rest is just a hallucination.

There is no extraction of actual visual memories from the individual.

I really hope that this won’t going to go any further it can be 1000 times worse than bite analysis if this ever spills into forensic science.

This is wild. I didn't know anyone was seriously working on this let alone it being this far along.