This is a short update on our progress in locating vectors in the latent space of the recently released stable diffusion model. Image memorability relates to the degree to whether a human recognizes the repetition of an image after a single view. Interpolation between memorability levels in stable diffusion will allow us to produce images of specific objects or scenes before adjusting their memorability to a specific level. The video below showcases some early results of raw interpolations between low and high memorability (no cherry-picking, quality of outcomes varies).
The results at this stage of the process is quite coarse and high-level, meaning that interpolations are performed between higher level semantic categories. The ultimate goal of this project is to be able to perform such interpolations within semantic categories.
Interpolation performed on random seeds from low to high memorability.
Using ViTMem for memorability estimation: https://pypi.org/project/vitmem/ https://github.com/brainpriority/vitmem