Flow-Augmented Variational Autoencoders: A Lightweight Alternative for Galaxy Distribution Generation

作者：Karunya Harikrishnan, Abhiroop I, Malathi M, K. G., D. S · 发表于：International Conference on Signal Processing, Communications and Computing · 年份：2025 · DOI：10.1109/icspcc66825.2025.11194375

Variational Autoencoders (VAEs) have emerged as a powerful generative model by learning efficient latent representations of data. However, their performance is often constrained by the simplistic assumption of a standard Gaussian prior, limiting their expressiveness and leading to suboptimal sample quality. In this work, the flexibility of the VAE’s is enhanced by approximating the VAE’s posterior by integrating Normalizing Flows (NF)—a series of invertible, learnable transformations that progressively map a simple distribution into a more complex one. By augmenting the encoder with NF-based transformations, we significantly improve the expressiveness of the latent space while retaining tractable inference and efficient sampling. Experiments conducted on the Galaxy Distribution data, known for its complicated distributions, demonstrate that the proposed VAE+NF architecture achieves lower negative log-likelihoods and generates higher-quality samples compared to baseline VAEs, thus effectively capturing multi-modal and non-linear dependencies. While the generative quality remains modest compared to state-of-the-art models like GANs or diffusion models, the approach offers notable gains in resource efficiency, making it an attractive tradeoff for scenarios where computational cost is a limiting factor.