Beauty is in the ELBO of the Beholder: A Variational Account of Processing Fluency in Face Perception

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Beauty is in the ELBO of the Beholder".

Jane: Here's the gist of the paper. The gist: Facial attractiveness can be explained in terms of a variational interpretation of processing fluency, where human attractiveness gradients closely match the evidence lower bound (ELBO) in ratedistortion space and independently learned latent spaces contain an attractiveness direction that transfers strongly across random initializations and training data.

Tom: Variational Interpretation of Processing Fluency The processing fluency theory posits that stimuli that are easier to encode, retrieve, or interpret tend to elicit more positive evaluations from human subjects <ref:2608.24219#pg2> In this view, preferences are not determined by the objective physical properties of a stimulus in isolation, but rather by how that stimulus “fits” into the observer’s perceptual system <ref:2608.24219#pg4>.

Jane: Probabilistic generative models provide a natural framework for modeling processing fluency <ref:2608.24219#pg5> Perception can be viewed as inference under an internal model of how sensory observations are generated <ref:2608.24219#pg6> VAEs make use of this principle explicitly by learning an approximate posterior qϕ(z x) together with a generative model pθ(x z) under a prior p(z) <ref:2608.24219#pg8> Because the exact marginal likelihood is generally intractable, VAEs minimize the negative evidence lower bound (ELBO) given by −ELBO(x) = Eqϕ(zx)

− log pθ(x z): distortion D(x) +β KL

qϕ(z x)∥ p(z): <ref:2608.24219#pg9> This decomposition suggests a concrete variational interpretation of processing fluency <ref:2608.24219#pg10> A stimulus is represented with ease when it can be explained with high fidelity (low distortion) and at a low representational cost (rate) <ref:2608.24219#pg10>.

Tom: The ELBO combines these two desiderata into a single quantity and thus provides a scalar measure for processing fluency Under this interpretation, we predict that aesthetic pleasure stems from a high ELBO (low distortion and low rate) induced by the learned generative representation.

Paper summary: Tom: We're diving deeper into "Beauty is in the ELBO of the Beholder: A Variational Account of Processing Fluency in Face Perception." We saw that human attractiveness ratings align with the evidence lower bound, or ELBO, across various face datasets.

Jane: So what does that mean for us? It suggests that the mathematical measure of how well a generative model represents a face is actually what reflects our own aesthetic judgment >

Tom: Precisely. The paper argues that beauty is tied to the ease with which a face can be perceived, and in the context of VAEs, that ease is captured by minimizing the negative evidence lower bound in rate-distortion space >

Lu: The core thesis here is using variational autoencoders to model processing fluency, where fluency means low distortion and low rate when reconstructing an image from its latent code >

Meng: It’s about finding a balance: you want the reconstruction to be accurate, low distortion, but you also don't want the latent representation to be too complex or high rate >

Lalam: And they found that this ELBO score is what determines our preference because it combines those two things into one scalar measure of fluency >

Jane: This is a big step because it takes the abstract idea of aesthetic pleasure and makes it concrete using something we already use in modern generative AI: the VAE framework >

Tom: And they show that this connection holds even when you look at different demographic groups and how the models are trained, because that attractiveness direction is conserved across those variations >

Lu: That conservation of the latent attractiveness direction is key. It implies there’s a fundamental geometric structure in the face space that we value aesthetically, which the AI has learned implicitly >

Jane: So, when we think about why some faces just "pop" at us, maybe it’s because their geometry fits into this specific low-distortion, low-rate manifold that the VAE learns to represent well >

Conclusion: Tom: So, wrapping up "Beauty is in the ELBO of the Beholder: A Variational Account of Processing Fluency in Face Perception," we’ve established that human attractiveness gradients match the ELBO direction across datasets >

Jane: And that this latent direction holds up even when models are trained on different data because it’s a property of how VAEs learn representations, not just one specific face set >

Tom: The authors found attractive faces are more prototypical in both shape and latent space, which means they are geometrically closer to the center of what we find appealing >

Lu: What this implies for the field is that we can use these learned geometric directions to guide our own models toward generating representations that humans find more appealing without needing manual attractiveness supervision >

Meng: For deployment, if a model learns this latent attractiveness direction, it could be used as a simple prior constraint to keep generated faces within a pleasing region of the data distribution >

Lalam: It suggests that aesthetic judgment isn't just subjective; it’s tied to the measurable geometry of the data itself when we use probabilistic models >

Jane: It shifts how we think about generative models from just creating realistic images to creating representations that are also perceptually favorable >