Mapping to new embedding spaces or using ensembles
PaulScotti opened this issue · 0 comments
PaulScotti commented
MindEye can map to any multimodal latent space like CLIP, OpenCLIP, AlexNet, ImageBind, etc. Can you improve our retrieval and/or reconstruction evaluation metrics by either (1) mapping to a different embedding space or (2) combining multiple models mapped to different embedding spaces to improve final metrics?