I'm currently working on tools for better latent space exploration in generative models. The latent space of Stable Diffusion for example, is incredibly rich, and the traditional txt2img and img2img pipelines for accessing it are only scratching the surface of what's possible. We're talking about billions of images compressed (lossily) into 4 GB — yet we're still using very basic interfaces for interacting with this neurally compressed data structure. I'm working on creating other ways to explore latent spaces like this. My current project is just to dip my toes in the water, and it's an animated music video generator built with Stable Diffusion. I hope to show it off here on HN in the near future.
I'm no expert, but this paper[0] has piqued my interest recently. From what I can kind of piece together, it discusses treating the latent space as a manifold from differential geometry, and creates a Riemannian metric for measuring distance between points on that manifold in such a way that the "curvature" of the manifold represents the semantic density of the latent space. This makes traversing between two points in a smooth/logical fashion easy by following the geodesics of the manifold. I think. Maybe. I understand maybe about 10% of what I just said to be honest, and still have lots to learn, which is part of why I find it so exciting.
