Ludovica Schaerf
On latent representations of vision generative models
In this thesis, I reconceptualize “latent representations” in generative vision models as technical abstractions and cultural constructs. Models are viewed here as cultural snapshots resulting from the combination of model choices and data: the weights sediment an episteme; the data insert biases. Representation, in this perspective, is not a neutral description but a historically situated “way of seeing”: representations are imbued with cultural canons, aesthetic hierarchies, and exclusions. It is more precisely a pre-presentation of what the generative models will generate.
Digital Visual Studies Webpage ( Ludovica Schaerf)