Stable "Retrieval-Augmented" Diffusion (WIP)
Update: Didn't work - moving on.
This is an experiment of mine using faiss indices built with clip-retrieval
This model was not originally trained with retrieval augmentation in mind, and so won't benefit as much from this.
Further, the latents produced the stable-diffusion autoencoder are hardly small. Storing more than ~100,000 can be a hassle.