Geometry-Preserving Loss Functions Enable Blackbox GAN Adaptation
A new arXiv preprint (2604.23888) proposes a method for adapting blackbox generative models without access to weights or gradients. The approach uses geometry-preserving loss functions with pre-trained GANs and rethinks GAN inversion for domain adaptation. This addresses challenges in fine-tuning large-scale generative AI tools that are not widely available due to storage costs and restricted access.
Key facts
- arXiv preprint 2604.23888 proposes geometry-preserving loss functions for blackbox generative model adaptation.
- Method uses pre-trained GANs and rethinks GAN inversion for domain adaptation.
- Traditional fine-tuning is infeasible due to storage costs and restricted access to weights and gradients.
- The approach is an end-to-end pipeline for domain adaptation.
- Published on arXiv with Announce Type cross.
Entities
Institutions
- arXiv