Pré-Publication, Document De Travail Année : 2024

Agnostic latent diversity enhancement in generative modeling

Résumé

Generative modeling methods can generate images from textual or visual inputs. However, diversity in the generated images persists as a major challenge of the existing approaches. In this work, we address this issue head-on by demonstrating that: (a) the diversity of a generated batch of images is intrinsically linked to the diversity within the latent variables; (b) leveraging the geometry of the latent space, we can establish an effective metric for quantifying diversity; and (c) employing this insight allows one to achieve a significantly enhanced diversity in image generation beyond the capabilities of traditional random independent sampling. This advancement is consistent across a variety of generative models, including Generative Adversarial Networks (GANs) and latent diffusion models. To facilitate further research and application in this field, we are also releasing a comprehensive package that enables easy reproduction of our experiments. We integrate our contributions into a widely recognized tool for generative image modeling, ensuring that our improvements are accessible to the broader community.

Fichier principal
Vignette du fichier
a_few_points_in_huge_hyperballs____new (35).pdf (4.99 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04661473 , version 1 (24-07-2024)
hal-04661473 , version 2 (28-10-2024)

Licence

Identifiants

  • HAL Id : hal-04661473 , version 2

Citer

Mariia Zameshina, Mathurin Videau, Alessandro Leite, Marc Schoenauer, Laurent Najman, et al.. Agnostic latent diversity enhancement in generative modeling. 2024. ⟨hal-04661473v2⟩
930 Consultations
1178 Téléchargements

Partager

  • More