Generative AI in Depth: A Survey of Recent Advances, Model Variants, and Real-World Applications

Shamim Yazdani 1 , Akansha Singh 2 , Nripsuta Saxena 3 , Zichong Wang 1 , Avash Palikhe 1 , Deng Pan 1 , Umapada Pal 4 , Jie Yang 5, Wenbin Zhang · 2025

Read Generative AI in Depth: A Survey of Recent Advances, Model Variants, and Real-World Applications by Shamim Yazdani 1 , Akansha Singh 2 , Nripsuta Saxena 3 , Zichong Wang 1 , Avash Palikhe 1 , Deng Pan 1 , Umapada Pal 4 , Jie Yang 5, Wenbin Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-21887.txt.

Introduction

Shamim Yazdani leads a nine-author survey of generative artificial intelligence, the model family behind synthetic images, video, and text. With research, uses, and open problems all growing fast, the authors set out to organize the field so newcomers and practitioners can see how the pieces fit together.

The survey builds a taxonomy covering the three main model families, adversarial networks, variational autoencoders, and diffusion models, along with their many variants and hybrids. It traces the innovations that improved output quality, variety, and control, and catalogs real-world uses across domains. Ethical concerns raised by systems that can fabricate convincing content receive direct attention as well. The result is a map of a fast-moving area, including the technical challenges still unresolved.

Catalog record

CREATORShamim Yazdani 1 , Akansha Singh 2 , Nripsuta Saxena 3 , Zichong Wang 1 , Avash Palikhe 1 , Deng Pan 1 , Umapada Pal 4 , Jie Yang 5, Wenbin Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE121 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-21887

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Generative AI in Depth: A Survey of Recent Advances, Model Variants, and Real-World Applications Shamim Yazdani 1 , Akansha Singh 2 , Nripsuta Saxena 3 , Zichong Wang 1 , Avash Palikhe 1 , Deng Pan 1 , Umapada Pal 4 , Jie Yang 5, Wenbin Zhang Abstract In recent years, deep learning based generative models, particularly Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Diffusion Models (DMs), have been instrumental in in generating diverse, high-quality content across various domains, such as image and video synthesis. This capability has led to widespread adoption of these models and has captured strong public interest. As they continue to advance at a rapid pace, the growing volume of research, expanding application areas, and unresolved technical challenges make it increasingly difficult to stay current. To address this need, this survey introduces a comprehensive taxonomy that organizes the literature and provides a cohesive framework for understanding the development of GANs, VAEs, and DMs, including their many variants and combined approaches. We highlight key innovations that have improved the quality, diversity, and controllability of generated outputs, reflecting the expanding potential of generative artificial intelligence. In addition to summarizing technical progress, we examine rising ethical concerns, including the risks of misuse and the broader societal impact of synthetic media. Finally, we outline persistent challenges and propose future research directions, offering a structured and forward looking perspective for researchers in this fast evolving field. keywords Generative Adversarial Networks, Variational Autoencoders, Diffusion Models, Artificial Intelligence, Deep Learning, generative AI. 1 Introduction The field of generative artificial intelligence (AI) has made significant strides within the swiftly evolving landscape of AI Gozalo-Brizuela and Garrido-Merchán 2023 ; Liu et al. 2021 ; Jin et al. 2023 ; Amon et al. 2025 . The capability of generative AI models to produce realistic images and videos has clear and noteworthy practical implications across a multitude of domains, such as art Elgammal et al. 2017 ; Li et al. 2017 , entertainment Ehtesham et al. 2025 , gaming Liu et al. 2021 , design Gatys et al. 2016 , medicine Pinaya et al. 2023 ; Jing et al. 2023 ; Shokrollahi et al. 2023 ; Ktena et al. 2023 , and data augmentation Perez and Wang 2017 for machine learning tasks. Leading the revolution in generative AI are the continuous surge in computing power and the emergence of innovative deep learning (DL) architectures such as Generative Adversarial Networks (GANs) Goodfellow et al. 2014 , Variational Autoencoders (VAEs) Kingma and Welling 2013 ; Rezende et al. 2014 , and Diffusion Models (DMs) Karras et al. 2023 ; Wu et al. 2023 ; Yang et al. 2023 ; Croitoru et al. 2023 ; De Bortoli et al. 2021 , which excel at capturing intricate data patterns and producing more realistic and varied images, surpassing traditional generative methods Bengesi et al. 2023 . Although just a decade old, these architectures have transformed the realm of generative AI in various domains. Introduced in 2014, GANs transformed the training of generative models with a new revolutionary framework consisting of two essential components: a generator and a discriminator. Locked in an adversarial competition Goodfellow et al. 2014 , the generator produces new samples, while the discriminator aims to distinguish between fake and real images. By incorporating an adversarial competition between the generator and discriminator, GANs can produce remarkably realistic and diverse image and video samples Ali et al. 2021 . GANs have found applications in various areas within the realm of visual media generation, including but not limited to image-to-image translation Isola et al. 2017 , object detection Zhang et al. 2020 , artistic style transferring Xu et al. 2021 , and even gesture generation in robots Nishimura et al. 2020 . On the other hand, VAEs leverage the capabilities of both autoencoders and probabilistic modeling to acquire insightful data representations within a latent space. In contrast to developing an encoder that generates a solitary value for each latent attribute, the VAE’s encoder is crafted to articulate a probability distribution for each latent attribute Kingma and Welling 2013 . This design facilitates seamless interpolation, meaningful sampling, and the generation of high-quality images. VAEs find application in image and video generation Gregor et al. 2015 ; Yan et al. 2021 , classification Gregor et al. 2015 , and environmental simulation Cheng et al. 2023 . The most recent among the three architectures, DMs have revolutionized data-driven image and video synthesis, establishing themselves as leaders in managing large datasets Karras et al. 2023 , evidenced by their remarkable proficiency in capturing intricate data distributions Wu et al. 2023 . DMs consist of two key phases: a forward diffusion stage and a reverse diffusion stage. The input data undergoes gradual perturbation with Gaussian noise across multiple steps in the forward phase, and the model systematically reconstructs the original input by reversing the diffusion process in the reverse stage. Artistic painting and text-guided image editing Croitoru et al. 2023 , MRI reconstruction Güngör et al. 2023 , self-driving cars Hu et al. 2023 , and video production from text Croitoru et al. 2023 are among the successful applications of DMs in the domain of image and video generation. Despite their transformative impact on generative AI, models such as GANs, VAEs, and DMs continue to face notable limitations. For instance, GANs demand meticulous stabilization during training Goodfellow et al. 2014 ; Arjovsky et al. 2017 , as the adversarial nature of the generator and discriminator optimized via alternating or simultaneous gradient descent can easily lead to instability, mode collapse, or non-convergence. VAEs, while offering a principled probabilistic framework, often produce blurry outputs when modeling complex data distributions, due to their reliance on pixel-wise reconstruction losses Dosovitskiy and Brox 2016 ; Larsen et al. 2016 . In contrast, Diffusion Models achieve impressive sample quality but at the cost of computational inefficiency, requiring hundreds to thousands of denoising steps to generate a single sample. To overcome these challenges, a wide range of strategies have been developed to improve training stability, enhance sample quality and diversity, and reduce inference time Arjovsky et al. 2017 ; Radford et al. 2015 ; Cheng et al. 2023 ; Isola et al. 2017 ; Zhao et al. 2017 ; Cao et al. 2022 ; Yang et al. 2023 . The resulting surge in research has significantly expanded the field, making it increasingly difficult to stay current with ongoing advancements and unresolved questions. To this end, this paper presents a comprehensive survey systematically covering the advancements in generative AI architectures as well as improvements and limitations. Specifically, this work, to the best of our knowledge, presents the first thorough study of the foundational breakthroughs, continued advancements, and current shortcomings in generative AI explicitly focusing on GANs, VAEs and DMs, for researchers and practitioners to have a thorough understanding of the current landscape. The primary contributions of this paper include: 1. We introduce a novel taxonomy of generative models, spanning GANs, VAEs, hybrid GAN-VAE architectures, and Diffusion Models (DMs) that categorizes them based on key design principles, offering a structured lens through which to understand the evolution of the field. 2. Building on this taxonomy, we provide a systematic and comprehensive review of these models,

Read the complete text