Comparative Analysis of Transformer, Diffusion, and GAN-Based Generative Models
Keywords:
generative models, transformers, diffusion models, GANs, comparative evaluation, generation quality, inference efficiency, paradigm comparisonAbstract
The generative AI landscape is currently defined by three dominant paradigmatic approaches -- autoregressive transformer models, diffusion models, and generative adversarial networks -- each with distinct training objectives, architectural principles, and generation characteristics that make them differentially suitable for different generation tasks, data modalities, and deployment contexts. Despite the substantial literature on each paradigm individually, a rigorous cross-paradigm comparative evaluation that assesses all three approaches on a common set of benchmarks, quality dimensions, and efficiency criteria remains absent, making it difficult for practitioners to make principled architecture selection decisions. This paper presents the first comprehensive cross-paradigm comparative evaluation of transformer, diffusion, and GAN-based generative models across eight generation domains -- image synthesis, text generation, audio synthesis, video synthesis, molecular generation, code synthesis, 3D object generation, and time-series forecasting -- using a unified evaluation framework covering six quality dimensions: sample quality, diversity, training stability, inference efficiency, controllability, and compositional generalisation. Results across 24 benchmark comparisons reveal that no single paradigm dominates across all domains and dimensions: diffusion models achieve the highest image and audio sample quality (FID = 14.2 for image, MOS = 4.6 for audio); autoregressive transformers achieve superior text quality (perplexity), code synthesis (HumanEval pass@1 = 52.4%), and compositional generalisation (42.6% C-CME R@1); GANs achieve the highest inference speed (6.8x faster than diffusion, 3.2x faster than transformers) at competitive image quality. The paper contributes a cross-paradigm evaluation framework, a domain-paradigm suitability matrix, and empirical evidence supporting a complementarity rather than competition view of generative AI paradigms.
