Reproduce “AVAE: Adversarial Variational Auto Encoder” on OpenResearch

About this paper

Antoine Plumerault, Hervé Le Borgne, Céline Hudelot

The Adversarial Variational Auto Encoder (AVAE) framework, developed by researchers at CEA List and Centrale-Supelec, addresses the long-standing trade-off between reconstruction fidelity and image realism in generative models. It generates sharp, photorealistic image reconstructions by treating them as plausible samples from the data manifold rather than pixel-wise averages, achieving competitive FID scores while maintaining low reconstruction error across multiple datasets.

Key insights

• Realistic image reconstructions should be plausible samples from the data manifold rather than pixel-wise averages, which often lead to blurriness.
• A generator can serve a dual role, acting as both a GAN generator for new images and a decoder for reconstruction tasks, utilizing an additional noise component during reconstruction to make plausible choices.
• The combination of a Manifold Loss (ensuring realism) and a Latent Space Loss (maintaining latent consistency) allows for simultaneous optimization of both reconstruction fidelity and image realism.

Summary by alphaXiv.