Eigenvector Fields for Latent Space Discovery

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autore principale: SÉRGIO DE ANDRADE, PAULO
Natura: Recurso digital
Pubblicazione: Zenodo 2025
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866901476365828096
author SÉRGIO DE ANDRADE, PAULO
author_facet SÉRGIO DE ANDRADE, PAULO
contents Deep generative models have demonstrated remarkable success in synthesizing high-fidelity data, yet understanding and controlling their internal latent spaces remains a significant challenge. This paper introduces a novel, unsupervised method for discovering meaningful and disentangled directions of variation within the latent space of pre-trained generative models. Our approach constructs a local affinity graph based on the semantic similarity of generated outputs for a neighborhood of latent space samples. By computing the eigenvectors of the graph Laplacian, we derive a set of orthogonal direction vectors that constitute a local vector field. These eigenvector fields correspond to the principal axes of semantic variation in the output space, providing an interpretable coordinate system for navigating the latent manifold. We demonstrate empirically, using a pre-trained StyleGAN model on the FFHQ dataset, that traversing the latent space along these eigenvector directions leads to coherent and disentangled semantic changes in the generated images, such as modifications to age, expression, and pose. Qualitative and quantitative results show that our method outperforms baseline approaches like Principal Component Analysis in identifying semantically rich and disentangled latent directions without requiring any external labels or classifiers. This work provides a principled geometric framework for enhancing the interpretability and controllability of deep generative models.
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_17682214
institution Zenodo
language
publishDate 2025
publisher Zenodo
record_format zenodo
spellingShingle Eigenvector Fields for Latent Space Discovery
SÉRGIO DE ANDRADE, PAULO
Deep generative models have demonstrated remarkable success in synthesizing high-fidelity data, yet understanding and controlling their internal latent spaces remains a significant challenge. This paper introduces a novel, unsupervised method for discovering meaningful and disentangled directions of variation within the latent space of pre-trained generative models. Our approach constructs a local affinity graph based on the semantic similarity of generated outputs for a neighborhood of latent space samples. By computing the eigenvectors of the graph Laplacian, we derive a set of orthogonal direction vectors that constitute a local vector field. These eigenvector fields correspond to the principal axes of semantic variation in the output space, providing an interpretable coordinate system for navigating the latent manifold. We demonstrate empirically, using a pre-trained StyleGAN model on the FFHQ dataset, that traversing the latent space along these eigenvector directions leads to coherent and disentangled semantic changes in the generated images, such as modifications to age, expression, and pose. Qualitative and quantitative results show that our method outperforms baseline approaches like Principal Component Analysis in identifying semantically rich and disentangled latent directions without requiring any external labels or classifiers. This work provides a principled geometric framework for enhancing the interpretability and controllability of deep generative models.
title Eigenvector Fields for Latent Space Discovery
url https://doi.org/10.5281/zenodo.17682214