The use of generative AI models increasingly involves the reliance on a few black box pretrained and centralized models where design intent is conveyed through language prompts. However, language is an imperfect medium for conveying design and artistic intent especially in the visual arts and architecture.
In this course we will explore some alternative generative ML methods and try to find the points in the process where design intent and experimentation can be injected. These points include the curation and procurement of the datasets, the training process, the assembly of models and most importantly the formulation of the ‘loss function’. The ‘loss function’ is the end point of any training or optimization process in ML that encodes what we want to model to ‘learn’. We will frame the ML mediated design process as an aesthetic optimization procedure where the loss function is the point where design intent is encoded.
The theme for this semester is ‘taste’ and we will focus on the training of models that can act as surrogates for personal aesthetic preferences. In a sense we want to create maximally biased and idiosyncratic models and then apply them to processes of curation, selection, and generation of visual artefacts. Students will be asked to collect and curate images that describe the personal artistic/aesthetic preferences of one or more people (e.g. artists or architects with a good record of photographic archives, writing and work) and use them to train models that will try to predict what that person would have found visually appealing. Then we will reverse the use of the models and try to ask how we can curate, modify and generate geometry and visual content that would maximize this appeal. The end goal is to design a series of artifacts (objects, furniture, paintings, postures) that are partly driven by the trained “surrogate of taste” model.
The class is structured as a small studio with a series of workshops introducing the tools necessary. No coding knowledge will be required.