Deep Generative Modelling: A Comparative Review of VAEs, GANs, Normalizing Flows, Energy-Based and Autoregressive Models

Analysis

Why This Paper Matters

Deep generative models have become a cornerstone of modern AI, enabling applications from image synthesis to data augmentation. However, the field has fragmented into multiple competing paradigms, each with its own strengths and weaknesses. This paper addresses a critical need for a comprehensive, accessible overview that clarifies the landscape. By systematically comparing VAEs, GANs, normalizing flows, energy-based models, and autoregressive models, it provides a roadmap for both newcomers and experienced researchers.

The paper's significance lies in its emphasis on the interconnections between these approaches. Rather than presenting them as isolated techniques, it highlights how they relate to each other, often through hybrid models. This holistic perspective is essential for advancing the field, as it encourages cross-pollination of ideas and helps researchers identify promising directions for innovation.

Technical Contributions

• Unified Framework: The paper organizes the five major generative model families under a common probabilistic framework, clarifying their mathematical relationships.
• Trade-off Analysis: It explicitly discusses the trade-offs each model makes in terms of run-time, sample diversity, and architectural constraints, providing a practical guide for model selection.
• Hybrid Approaches: The review covers numerous hybrid models that combine elements from different families, such as VAEs with GANs (VAE-GANs) and autoregressive models with normalizing flows.
• State-of-the-Art Review: It summarizes recent advances and implementations, giving readers a snapshot of the field's progress up to 2021.

Results

As a review paper, it does not introduce new experimental results. Instead, it synthesizes existing knowledge, offering qualitative comparisons. For example, it notes that GANs typically produce sharp samples but suffer from mode collapse, while VAEs offer stable training but produce blurrier outputs. Normalizing flows provide exact likelihood estimation but are computationally expensive. Autoregressive models achieve high-quality samples but are slow to generate. Energy-based models are flexible but challenging to train. These insights are drawn from the literature, not from new experiments.

Significance

The paper's impact is primarily educational and organizational. It serves as a comprehensive reference that can help researchers quickly understand the strengths and limitations of each approach, potentially guiding them toward more effective model design. By highlighting the interconnections between methods, it may inspire novel hybrid models that leverage the best of multiple paradigms. The paper's citation count (722) indicates its widespread use as a foundational resource in the deep generative modeling community. Its publication at Durham University adds to the academic credibility, and its open-access nature ensures broad accessibility. Overall, this review contributes to the maturation of the field by consolidating knowledge and fostering a more integrated understanding of deep generative models.