Hi Paul,
First of all, thank you for your fantastic work on MindEye and MindEye2—they’ve been incredibly insightful and have significantly advanced the fMRI-to-image field.
I noticed your suggestions in the MindEye repository issues regarding potential improvements to the low-level reconstruction pipeline, such as exploring different VAEs, novel training strategies, or even integrating ControlNet. At the same time, I saw that MindEye2 retained a relatively similar low-level reconstruction approach despite the major gains from ViT-bigG/14.
I’d love to hear your thoughts on this—were there specific considerations behind keeping the low-level pipeline largely unchanged? Or is there something I might have overlooked in the MindEye2 paper?
Thanks again for your brilliant contributions, and I really appreciate any insights you might share!
Hi Paul,
First of all, thank you for your fantastic work on MindEye and MindEye2—they’ve been incredibly insightful and have significantly advanced the fMRI-to-image field.
I noticed your suggestions in the MindEye repository issues regarding potential improvements to the low-level reconstruction pipeline, such as exploring different VAEs, novel training strategies, or even integrating ControlNet. At the same time, I saw that MindEye2 retained a relatively similar low-level reconstruction approach despite the major gains from ViT-bigG/14.
I’d love to hear your thoughts on this—were there specific considerations behind keeping the low-level pipeline largely unchanged? Or is there something I might have overlooked in the MindEye2 paper?
Thanks again for your brilliant contributions, and I really appreciate any insights you might share!