Stable Diffusion with Brain Activity

sites.google.com

2 points·aloukissas··5 comments

Already discussed, appears to be BS https://news.ycombinator.com/item?id=35012981

I think not. The top comment just claims the model was overtrained, which is true, but it was still picking images out of the modest dataset from the fMRI brain scans.

Thats still amazing! And reasonable, as the researchers can only scan so many brain-image pairs.

Point is the whole decoding and image bit is not real. Its matching brain activity to a pre-existing category with some extra smoke and mirrors on top about generating images. The visual aspect is absent, which degrades it from visually decoding our thoughts to showing an unrelated picture of the thing we're thinking about.

Oh so its like image to text, and then text to diffusion, but the text is just correlated to what they trained on. Yeah that is a can of nothing unless the scan to text is using CLIP interrogate or something.

That was roughly my take reading the discussion yesterday. It's possible there is something that's been missed in the criticism, but I haven't seen any compelling evidence that this is actually decoding what we visualize (in the sense of the actual visual) and so the SD angle is just a gag.

Problem is, people see the article, don't understand what it means, and start parroting it: https://news.ycombinator.com/item?id=35023642

Reply on news.ycombinator.com