I've tried to simplify this in interviews (& I suck at math): The deep learning process organizes all faces along a main face-axis in multidimensional space. If you follow this vector to its end you find Loab: the last face; an eigenface that one type of face-ness extends toward.
https://twitter.com/Elrohir_Ancalin/status/1567443408930672640
Probably, all the different face-nesses are represented on different axes and the Loab axis is just one of many. All faces are found in relationship to these axes. Collapse the idea to an x,y plane: imagine a simple system of 2 face-ness axes, where one is a Loabness axis.
The map of possible faces is finite though, so it has a geometry with edges and vertices and outward protrusions. A negative weighted prompt makes you likely to get caught inside one of these extremities with the spooky eigenfaces and the very few non-face things that are nearby.
This stuff is really weird and fun. VQGAN+CLIP Imagenet had EigenBirdBugBushes. Malevich had EigenCars.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
