[Paper Review] A Complete Survey on Generative AI (AIGC): Is ChatGPT from GPT-4 to GPT-5 All You Need?

Chaoning Zhang, Chenshuang Zhang+15 more|arXiv (Cornell University)|Mar 21, 2023

Artificial Intelligence in Healthcare and Education105 citations

TL;DR

This paper surveys generative AI (AIGC) fundamentals, techniques, tasks, applications, and challenges, examining how GPT-era models (GPT-4 and beyond) enable diverse content creation across text, image, video, and more.

ABSTRACT

As ChatGPT goes viral, generative AI (AIGC, a.k.a AI-generated content) has made headlines everywhere because of its ability to analyze and create text, images, and beyond. With such overwhelming media coverage, it is almost impossible for us to miss the opportunity to glimpse AIGC from a certain angle. In the era of AI transitioning from pure analysis to creation, it is worth noting that ChatGPT, with its most recent language model GPT-4, is just a tool out of numerous AIGC tasks. Impressed by the capability of the ChatGPT, many people are wondering about its limits: can GPT-5 (or other future GPT variants) help ChatGPT unify all AIGC tasks for diversified content creation? Toward answering this question, a comprehensive review of existing AIGC tasks is needed. As such, our work comes to fill this gap promptly by offering a first look at AIGC, ranging from its techniques to applications. Modern generative AI relies on various technical foundations, ranging from model architecture and self-supervised pretraining to generative modeling methods (like GAN and diffusion models). After introducing the fundamental techniques, this work focuses on the technological development of various AIGC tasks based on their output type, including text, images, videos, 3D content, etc., which depicts the full potential of ChatGPT's future. Moreover, we summarize their significant applications in some mainstream industries, such as education and creativity content. Finally, we discuss the challenges currently faced and present an outlook on how generative AI might evolve in the near future.

Motivation & Objective

• Explain the fundamental techniques underpinning AIGC, including backbone architectures and self-supervised pretraining.
• Review AIGC tasks by output type (text, image, video, 3D, etc.) and their technological progress.
• Summarize industry applications of AIGC in education, media, advertising, and creative fields.
• Discuss challenges, ethical considerations, and future outlook for generative AI.

Proposed method

• Classify AIGC techniques into two categories: creation techniques (GANs, diffusion models) and general techniques (Transformers, self-supervised pretraining).
• Describe backbone architectures (RNNs, Transformers, CNNs, ViT, Swin, DeiT, etc.) and their roles in AIGC.
• Summarize self-supervised pretraining methods for language and vision (e.g., BERT, GPT, MAE, CLIP) and cross-modal pretraining (CLIP, ALIGN).
• Explain likelihood-based vs. energy-based generative models and relate GANs/diffusion models to energy-based perspectives.

Experimental results

Research questions

• RQ1 What are the fundamental techniques enabling modern AIGC tasks?
• RQ2 How do base architectures and pretraining strategies support diverse AIGC outputs across modalities?
• RQ3 What is the landscape of AIGC tasks and applications, and how might future GPT variants influence them?
• RQ4 What challenges and societal implications arise from widespread AIGC deployment?

Key findings

• AIGC rests on two technique classes: creation models (GANs, diffusion) and general techniques (Transformers, self-supervised learning).
• Transformers and ViTs have become core backbones in NLP and CV, enabling scalable AIGC models.
• Self-supervised pretraining and cross-modal learning (e.g., CLIP, ALIGN) are crucial for large-scale AIGC capabilities across text and image tasks.
• AIGC tasks span text generation, image generation, and beyond (video, 3D, speech, graphs), with rapid progress in text-to-image and multimodal generation.
• The rise of AIGC tools is tied to data access and compute resources enabling large-scale models like GPT-4-era systems.