MindEye currently has a tendency to output overly saturated reconstructions. This might explain why our SSIM evals are weirdly bad compared to the rest of the metrics. Our simple fix was to change cfg from 7.5 to 3.5, which did help, but it's clearly not the ideal solution.
One suggestion to fix this problem is to implement dynamic cfg scaling:
MindEye currently has a tendency to output overly saturated reconstructions. This might explain why our SSIM evals are weirdly bad compared to the rest of the metrics. Our simple fix was to change cfg from 7.5 to 3.5, which did help, but it's clearly not the ideal solution.
One suggestion to fix this problem is to implement dynamic cfg scaling:
https://discord.com/channels/1025299671226265621/1062896575430533133/1098730791330123820
(my attempt to implement this is in utils.py via the dynamic_cfg function, which currently does not work)