This work explores the structural unconscious of generative AI models through the lens of 1,100 flowers produced by Stable Diffusion. Each image was segmented using SAM (Segment Anything Model) to extract the computational contour — the shape that the machine perceives as "flower."
On the left, a three-dimensional UMAP projection maps the latent space of the model: each point is a flower, positioned according to its hidden features. The camera follows a walking path through this space, tracing connections between neighboring images. When switching between modes — Color, Shape, Texture, Composition — the entire cloud reorganizes, revealing how different criteria restructure the same set of images.
On the right, the extracted contour morphs continuously from one flower to the next inside a fixed circle. Ghost traces of previous forms linger and dissolve, building a palimpsest of computational perception. What emerges is not the flower itself, but the model's idea of a flower — its latent archetype, perpetually transforming.
The proximity score measures the distance between consecutive images in each mode's feature space, making visible the varying degrees of similarity that the model perceives but never articulates.