For the post about Amnesty International's use of image editing, I agree that generative models shouldn't be used to edit images, especially for media that will be circulated in the news / in a legal setting. However, I do also believe that it is fair to use generative models to synthesize visualizations with strong disclaimers about its origin (i.e. the fact that it is synthetic + the method used to synthesize it), which is already common practice in the news (i.e. there's this company called SITU Research that uses photogrammetry techniques to create layouts / 3D reconstructions of events). Really interesting ethical debate though.
The fMRI work is super cool, there was also this CVPR paper that used Stable Diffusion to analyze brain signals: https://sites.google.com/view/stablediffusion-with-brain, where they map the fMRI signals from the early visual cortex to optimize the noise input and late visual cortex to optimize the text conditioning that is the input to the diffusion model.
Although fMRI as it stands is still a really noisy signal (and much harder to use compared to asking someone to sketch what they're thinking), so it's interesting to think about what is the best UI for humans to interact with generative models and if measuring tools for brain signals (or maybe the software for analyzing them) will become advanced enough to be useful.
For the post about Amnesty International's use of image editing, I agree that generative models shouldn't be used to edit images, especially for media that will be circulated in the news / in a legal setting. However, I do also believe that it is fair to use generative models to synthesize visualizations with strong disclaimers about its origin (i.e. the fact that it is synthetic + the method used to synthesize it), which is already common practice in the news (i.e. there's this company called SITU Research that uses photogrammetry techniques to create layouts / 3D reconstructions of events). Really interesting ethical debate though.
The fMRI work is super cool, there was also this CVPR paper that used Stable Diffusion to analyze brain signals: https://sites.google.com/view/stablediffusion-with-brain, where they map the fMRI signals from the early visual cortex to optimize the noise input and late visual cortex to optimize the text conditioning that is the input to the diffusion model.
Although fMRI as it stands is still a really noisy signal (and much harder to use compared to asking someone to sketch what they're thinking), so it's interesting to think about what is the best UI for humans to interact with generative models and if measuring tools for brain signals (or maybe the software for analyzing them) will become advanced enough to be useful.