Latent feature spaces

I spent part of my day generating simple icon images at the rate of about one every two seconds. I generate them in batches of about 12 where there is slight variation in the images which sometimes lets me see transitions in the latent feature space, like how a red and maroon gradient in one shape might transition into two shapes, one red and the other maroon.

In the first set of images the overall shape starts as an almond shape, but in other images the prompt generates a circle with leaf shapes, or a flower design. The image is traversing a latent feature space. It’s a multidimensional feature space, so you get other feature variations like the shift from monochrome to multicolor and back.

The next set of images show how “knowledge” of physics is present in the latent feature space. On the left the shape has a sturdy base which affords sitting on a flat surface. The figure on the right resembles a dreamcatcher design and affords hanging from a nail. I’ll see various tear drop shapes. but never with the point facing down. Heavy parts of a shape tend to be seen at the bottom of objects.