Source-linked AI summary
A Complete Survey on Generative AI (AIGC): Is ChatGPT from GPT-4 to GPT-5 All You Need?
Chaoning Zhang, Chenshuang Zhang, Sheng Zheng, Yu Qiao, Chenghao Li, Mengchun Zhang, Sumit Kumar Dam, Chu Myaet Thwal, Ye Lin Tun, Le Luang Huy, Donguk kim, Sung-Ho Bae, Lik-Hang Lee, Yang Yang, Heng Tao Shen, In So Kweon, Choong Seon Hong
TL;DR
The paper addresses whether future GPT variants could unify diverse AIGC tasks, motivating a comprehensive review. It surveys technical foundations, task developments, applications, challenges, and possible evolution, concluding with a content-centered account of generative AI across these dimensions.
Problem
The paper asks whether future GPT variants could help ChatGPT unify all AIGC tasks, motivating a comprehensive review of diverse content-generation tasks.
Method
The paper surveys generative AI through AIGC tasks, covering underlying techniques, task-wise technological development, industry applications, and social impact.
Results
The survey covers technical foundations, output types including text, images, video, and 3D, industry applications, current challenges, and an outlook on future evolution.
Takeaways & Limitations
ChatGPT is presented as one tool among numerous AIGC tasks, while the survey provides a broad view of the field's development and applications.
Takeaways & Limitations
The introduction notes that media coverage of AIGC tends to be biased or sometimes misleading, and that ethical concerns have emerged with real-world applications.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
This introduction distinguishes generative AI from conventional analytical AI, then surveys AIGC techniques, output types, applications, challenges, and future directions. It frames the survey around whether future GPT variants could unify diverse AIGC tasks.
- Concept and scope: Generative AI creates new content, whereas conventional AI primarily analyzes existing data through classification or regression.The boundary is not absolute because generative AI often first understands existing data, and analytical tasks can also generate low-dimensional labels.
- Concept and scope: AIGC denotes content-generation tasks, while generative AI additionally includes the technical foundations supporting those tasks.The paper uses AIGC as shorthand for content-generation tasks despite the terms' close relationship.
- Survey organization: The survey organizes underlying techniques into generative modeling methods such as GANs and diffusion models, plus backbone architectures and self-supervised pretraining.Transformer, BERT, and MAE are cited as examples of the second class, which supports competitive AIGC performance.
- Survey organization: AIGC tasks are categorized by output type, covering text and image generation alongside video, 3D, speech, and multimodal-input tasks.The survey highlights ChatBot and machine translation for text, and image restoration, editing, and text-to-image for images.
- Applications and outlook: The paper surveys AIGC applications in entertainment, digital art, media and advertising, and education, while also discussing ethical challenges and future evolution.It presents the survey as covering basic techniques, task-wise development, industry applications, and social impact.
2 OVERVIEW
Generative AI marks a shift from analyzing existing data toward creating new content, supported by growing data access, computing resources, and generative techniques. The overview traces rising public interest, evolving content production, and the expanding scope of AIGC tasks.
- Popularity indicated by search interest: Search interest in generative AI increased significantly during the past year, especially after October 2022, with a similar trend for AIGC.For both terms, major hot regions include Asia, Northern America, and Western Europe, while China ranks highest among countries.
- Popularity indicated by search interest: Search terminology varies geographically: China records a generative-AI-to-AIGC interest ratio of 15/85, whereas the US records 90/10.Russia and Brazil record ratios of 100/0, and the authors state that China’s preference for AIGC is not fully clear.
- Content need: AIGC addresses the volume–quality trade-off between professional and user-generated content by enabling users to create content from ordinary inputs such as text descriptions.Text-to-image tools let users produce images without the long practice previously needed for decent-quality drawing.
- Content need: AIGC broadens creation beyond images to other content types and makes the boundary between content consumers and creators more flexible.Users can generate original, tailored content with greater control over what they produce and consume.
3 FUNDAMENTAL TECHNIQUES BEHIND AIGC
The paper organizes AIGC foundations into general techniques, which support development without directly generating content, and creation techniques, which generate content. It reviews backbone architectures and representative methods across NLP and computer vision.
- AIGC fundamentals are divided into general techniques and creation techniques.Creation techniques directly generate content, whereas general techniques such as Transformer architectures are essential for AIGC development but do not generate content directly.
- Deep learning advances AIGC through data-driven parameter optimization, improved backbone architectures, and larger datasets.
- Backbone architecture: Transformers replace recurrence with self-attention, using encoder-decoder components, multi-head attention, feed-forward networks, residual connections, and layer normalization.They construct global dependencies but require positional encoding to represent input positions.
- Backbone architecture: RNNs process sequential data through recurrent hidden states, while LSTM and GRU address long-term dependencies with added gating or simplified state mechanisms.LSTM uses a cell state and three gates but has higher complexity and memory requirements; GRU merges cell and hidden states and simplifies the gates.
- Backbone architecture: CNNs extract image features with shared convolution kernels, while variants improve diversity, trainability, connectivity, or efficiency through architectural changes.Examples include GoogleNet, ResNet, DenseNet, and EfficientNet.
- Backbone architecture: ViT converts images into patch-token sequences, adds a class token and positional embeddings, and feeds them into a standard Transformer.Swin reduces complexity through local-window attention, while DeiT reduces reliance on large datasets through teacher-student distillation.
4 AIGC TASK: TEXT GENERATION
Text generation spans translation, summarization, dialogue, and captioning, with controllability and multimodality as important directions. The survey covers modular and end-to-end dialogue systems alongside autoregressive and pretrained approaches for language generation.
- Text generation transforms text inputs into outputs for tasks including machine translation, summarization, dialogue, and image captioning.
- Dialogue systems: Dialogue systems include task-oriented systems for specific goals and open-domain systems without task or domain restrictions.
- Dialogue systems: Task-oriented dialogue pipelines commonly use natural language understanding, dialogue state tracking, dialogue policy learning, and natural language generation.
- Dialogue systems: Modular dialogue systems face coordination limits because improving one module may not improve the whole system, while end-to-end systems still face data and quality challenges.
- ChatGPT achieved unprecedented success in open-domain dialogue and performed competitively with commercial machine-translation products such as Google Translate.
- Machine translation: Neural machine translation progressed from recurrent and convolutional architectures toward Transformers, while low-resource settings motivate multilingual and pivot-translation strategies.
5 AIGC TASK: IMAGE GENERATION
Image generation includes restoration, semantic editing, and text-to-image synthesis, with image-type and text-guided controls. Diffusion models increasingly address the limitations of earlier generative approaches and support broader editing capabilities.
- Image synthesis is organized by input control, including image-conditioned restoration, editing, translation, and text-guided generation.
- Image restoration: Image restoration maps degraded images to clean images, but the inverse problem is ill-posed because a degraded input can have infinitely many clean-image mappings.
- Image restoration: GANs and diffusion models can generate multiple clean-image variations from one degraded image, while GANs commonly face complex training and mode collapse.
- Image restoration: Multi-task restoration seeks to replace separate models for each degradation with a single network capable of handling multiple degradations.
- Image editing: Image editing changes image semantics through content, style, or object-attribute manipulation, including style transfer and facial attribute editing.
- Text-to-image generation: Text-guided image generation produces images from textual descriptions, while diffusion-based editing aligns target images with text and can automatically predict editable masks.
6 AIGC TASK: BEYOND TEXT AND IMAGE
Beyond text and images, generative AI covers video, 3D objects, speech, graphs, music, and lip-reading. These modalities introduce challenges involving temporal coherence, scarce data, representation choices, stability, and computational cost.
- Video generation: Earlier video-generation systems were limited to short videos, simple scenes, and low-resolution datasets.
- Video generation: Video generation is harder than image generation because models must generate pixels while maintaining semantic coherence across frames.
- Video generation: Diffusion models advanced text-guided video generation, including approaches that extend text-to-image models and reduce reliance on paired text-video data.
- 3D generation: 3D content can use depth images, voxel grids, point clouds, meshes, or neural fields, each offering different advantages and disadvantages.
- 3D generation: 3D generation supports text, image, and 3D-data inputs, but remains challenging because of limited 3D data and unsuitable architectures.
- Speech synthesis: Diffusion models are increasingly used for speech synthesis because they avoid discriminators and provide more stable, simpler training than GANs.
- Graph generation: Graph-generation methods include realistic generation and goal-directed optimization, while traditional, autoregressive, and GAN-based approaches face structural or training limitations.
7 INDUSTRY APPLICATIONS
The survey highlights education as a major application area for generative AI, including personalized learning, tutoring, course-material creation, assessment, and student writing support. It also emphasizes risks involving incorrect information, bias, privacy, and the need for careful validation.
- Education: Generative AI can support education through personalized tutoring, course-material creation, assessment, evaluation, and interactive learning.
- Education: ChatGPT may reduce instructors’ workload, freeing time for academic productivity or developing more complex teaching skills.
- Education: Educational use requires careful evaluation and validation because generative AI may produce incorrect or biased material and raise privacy concerns.
- Education: ChatGPT can help students write essays, complete at-home tests or quizzes, understand concepts, and improve writing through feedback.
- Education: Generated course materials can be tailored to students’ learning styles and pace, potentially improving engagement and learning outcomes.
7.2 Game and metaverse
Generative AI personalizes games and virtual worlds by adapting avatars, scenarios, storylines, and interactive events, while expanding media accessibility and reporting formats.
- Generative AI enables users to customize avatars while providing diverse scenarios and storylines that make game and metaverse experiences more immersive.
- GPT-3-powered AI Dungeon generates open-ended stories and new events in response to users’ actions, creating unexpected gameplay.
- Games and metaverse platforms give users greater control over virtual experiences, including designing unique avatars and tailoring content to individual needs.
- Generative AI supports diverse media applications, including writing robots, AI news anchors, caption generation, and multilingual or accessible information services.
- The Associated Press generates roughly 40000 stories annually, while Quakebot completes a related earthquake article in three minutes.
- AI-driven sign-language and subtitle services helped people with hearing impairments access Beijing Winter Olympics broadcasts more immersively.
7.4 Advertising
AIGC gives advertising tools for generating personalized text, posters, and logos, helping advertisers create engaging content at scale while saving time and resources.
- SGS-PAC automatically personalizes advertising copy to meet individual consumer needs, addressing exposure to identical content.
- Vinci generates advertising posters from user-specified product images and slogans using deep-generation models.
- Brandmark.io automatically produces multiple logo variations from user preferences and specifications for business use.
- Keyword-constrained generation automates product listings intended to attract potential buyers on peer-to-peer marketplaces.
- AIGC enables advertisers to create highly personalized and engaging content at scale while saving time and resources.
7.5 Movie
AI is integrated across movie production, supporting scripts, visual effects, subtitles, speech synthesis, color grading, and editing, while enabling precise but costly character transformation.
- AI supports movie creation through script generation and revision, visual effects, sound effects, editing, labeling, video retrieval, and viewing platforms.
- VFX workflows use captured facial data, high-resolution 3D models, and AI manipulation to create convincing age transformations.
- Comparable facial-data and motion-capture workflows were used for age reduction in Blade Runner 2049 and Gemini Man.
- Deepfake and VFX techniques can recreate lifelike appearances of deceased actors, as illustrated by Paul Walker’s posthumous appearance in Fast and Furious 7.
- Automated SDH subtitles transcribe speech, speaker changes, and background noise, while NLP supports multilingual subtitles.
- AI-generated speech synthesis can narrate silent films and dub movies into multiple languages using systems trained on large human-audio datasets.
7.6 Music
AIGC is reshaping music and visual art by assisting composition, lyric writing, audio processing, painting, restoration, and personalized creative work.
- AI helps musicians identify musical patterns, improve digital audio, distribute music, overcome writer’s block, and generate lyrics, melodies, and soundtracks.
- AI-generated music tools include Sony CSL Flow Machine, AIVA, Amadeus Code, Amper, Ecrett Music, Magenta, and Watson Beat.
- AIVA lets users select a style and modify musical variables before recombining learned characteristics into new works.
- Generative AI produces diverse art forms, including abstract, Chinese shanshui, and Chinese ink paintings, and supports artwork conservation and restoration.
- MidJourney’s Space Opera Theatre won first place in the Colorado State Fair Art Competition, demonstrating the capability of AI painting tools.
- ControlNet adds drawing, depth, line-map, normal-map, and pose constraints to diffusion models, increasing generated-image variability.
- AI can create collaborative and customized artwork by combining artists’ styles or adapting outputs to users’ preferences and interests.
7.8 Code development
AIGC supports software development through code generation, review, refactoring, testing, and dataset augmentation, while also enabling creative image, video, voice, and avatar applications.
- Code development: AIGC can generate code, assist code review and refactoring, and help developers improve code quality without changing functionality.GitHub Copilot supports automatic code completion, while DeepCode provides suggestions for code improvement.
- Code development: Automatically generating high-quality test cases can improve testing efficiency and increase test coverage.
- Code development: Neural-network-based dataset augmentation can improve data-engineering efficiency, while explainability remains an interpretability requirement for generated code.
- Creative applications: FaceApp illustrates the reach of fun-oriented AIGC applications, with over 500 million downloads reported.
- Creative applications: AIGC applications include image and video editing, face swapping, and digital avatars that can track facial movements and resemble users.Digital-avatar systems evolved from preset characters to personalized customization and became standard smartphone features.
- Creative applications: Voice-changing tools modify pitch, timbre, and speech rate for entertainment, live streaming, and other communication purposes.
7.10 Other fields
Beyond mainstream content creation, AIGC is presented as useful for drug discovery, manufacturing design, medical consultation, and other professional workflows.
- Drug discovery: Drug discovery motivates AIGC because developing a novel drug averages around $3 billion and more than 10 years before market acceptance.
- Drug discovery: AlphaFold predicts protein structures, while AlphaFold2 improves on its predecessor and ProteinMPNN rapidly designs task-specific protein sequences.ProteinMPNN can generate entirely new proteins in a few seconds.
- Other applications: AIGC can improve worker efficiency through applications such as chatbots for basic medical advice and computer-aided manufacturing design.In medical consultation, severe cases can be referred to doctors; in manufacturing, AIGC can reduce repetitive design effort.
8 CHALLENGES AND OUTLOOK
The survey identifies interpretability, ethical and legal risks, domain-specific reliability, and practical deployment constraints as major AIGC challenges, while projecting more controllable models, finetuning, and broader startup participation.
- Challenges: AIGC models require substantial training data and compute resources, and their real-world applications still face numerous challenges.
- Challenges: Limited interpretability makes it difficult to understand, control, and correct undesirable model outputs.
- Challenges: AIGC raises concerns about data bias, copyright infringement, privacy violations, cheating, and misleading political content.
- Challenges: Domain-specific models remain necessary because systems such as Stable Diffusion and ChatBot can produce undesired images or factual mistakes.
- Outlook: Text instruction has improved image-generation control over early GAN-based approaches, but finer-grained control remains necessary.
- Outlook: Future development is expected to shift from pretraining toward finetuning, with more startups emerging as demand increases.Finetuning must balance foundation-model general capability against adaptation performance on new tasks.