AI Is Redesigning Fashion: What 163 Studies Reveal About Generative Design’s Rise and Risks

Artificial intelligence that can generate images, text, and designs has moved from laboratory curiosity to a working tool across the fashion industry, and researchers have now mapped exactly how that happened. A systematic review published in the journal Fashion and Textiles analyzed 163 publications drawn from four major databases—Web of Science, EBSCO, IEEE Xplore, and Scopus—to chart the research landscape of AI-generated content, or AIGC, in fashion. The review, conducted by Rufan Lin, Yihan Yu, Fan Xia, and Manhao Guan and published on 13 July 2026, finds that scholarship on the topic has grown explosively since 2022, that the underlying technology has shifted decisively from adversarial networks toward diffusion models and large language models, and that the field’s benefits and dangers are tightly intertwined at every level of the industry.

The scale of the growth is striking. Publications on AIGC in fashion began to increase significantly in 2022 and peaked in 2025 within the review’s dataset, which collected literature through December 20, 2025. The authors link this inflection directly to a technological breakthrough: the maturation of diffusion models around 2022 dramatically lowered the barrier to image generation and design applications, triggering a wave of application-layer research. Because database indexing lags mean some 2025 studies were likely missed, the authors note the true total for that year is probably higher than their statistics show, and they predict the upward trend will continue into 2026. Geographically, the reviewed literature shows an international pattern the authors describe as multi-center distribution with broad participation, with China, South Korea, and the United States producing the highest publication volumes, and countries including India, Italy, the United Kingdom, Malaysia, and Turkey also actively contributing.

The review followed PRISMA guidelines, the standardized framework for systematic literature reviews. The team searched the four databases with queries combining artificial intelligence terms such as machine learning, deep learning, and generative design with fashion terms such as apparel design and fashion technology, covering January 2015 through December 2025. After three weeks of screening titles and abstracts and removing duplicates, 183 articles remained; two further weeks of full-text reading narrowed the corpus to 163. Of these, 141 were original articles and 22 were reviews. The authors then applied a structured coding scheme and, for the qualitative portions of the analysis, used grounded-theory-style open, axial, and selective coding supported by NVivo 11 software.

One of the review’s clearest findings concerns how the technology itself has evolved. The authors divide this evolution into three stages. In the preliminary stage, fashion AI relied on discriminative models—convolutional neural networks that could extract silhouette structures, texture details, and styling features from clothing images, and recurrent neural networks that excelled at modeling sequential data such as consumer preferences and sales trends. These systems were analytical observers rather than creators: their outputs were numerical indicators, trend curves, or static reports based on existing data, and they did not directly participate in content generation or consumer interaction.

The second stage arrived with generative adversarial networks, or GANs, which the review identifies as the first wave of automated visual generation in fashion. Through adversarial training between a generator and a discriminator, GANs learned to synthesize fashion images with structural consistency, enabling breakthroughs in virtual try-on and style transfer. Yet the review stresses that this generation remained essentially reconstruction and transfer of existing visual data rather than original design driven by semantic intent. GANs also carry a known technical weakness: mode collapse, a failure mode in which the generator produces limited varieties of output despite high efficiency.

The third and current stage is built on diffusion models, which generate images through iterative denoising and, according to the review, offer superior stability compared with adversarial training. These models enable zero-shot generation from natural language prompts, producing novel designs without retraining for specific styles. In high-fidelity applications such as virtual try-on, diffusion models are gradually replacing GANs as the mainstream approach; the review highlights IDM-VTON, a system that fuses clothing semantic features into the cross-attention and self-attention layers of a diffusion model to achieve realistic virtual try-on in uncontrolled, real-world scenarios. Platforms such as Stable Diffusion and Midjourney have pushed this paradigm into design practice. More recently, researchers have begun building fashion agents that use large language models as core decision-making units—for example, the DressCode system, which allows 3D clothing synthesis and editing through natural language conversation. The authors characterize this stage as a shift from market prediction to market participation.

Thematically, the 163 publications cluster into five areas: intelligent design and customization, trend forecasting and consumer behavior, AI and fashion education, law and ethics, and sustainable fashion design. Intelligent design and customization is the largest theme, spanning algorithm optimization, semantic understanding, visual perception, and consumer co-creation. Representative work includes GAN frameworks for outfit coordination and color-texture separation, diffusion-based systems for converting designer sketches into garments, hybrid ResNet-Transformer and Swin Transformer architectures for classifying women’s shirts and trousers, and chatbots evolving from customer service tools into co-creation partners that let users modify garments through semantic instructions. In trend forecasting, reviewed studies include a neuro-fuzzy inference system analyzing Internet of Things data for real-time market prediction and a forecasting system integrating computer vision and machine learning to identify trends in color, style, and garment combinations.

The dominant research paradigms are empirical. Among the 141 original articles, case-based and experimental research and quantitative research dominate, followed by mixed, theoretical, and qualitative approaches—a pattern the authors connect to the field’s focus on building and validating generative algorithms, which requires quantitative performance benchmarks and concrete design cases. The literature is also widely scattered across interdisciplinary journals: sustainability journals, textile and clothing journals including Fashion and Textiles itself, engineering and computer science venues focused on algorithms, multimedia journals focused on visual generation, and human-computer interaction journals focused on user experience.

The review’s central conceptual contribution is a three-layer framework—application scenario, design model, and attribute—built from qualitative coding of the literature’s stated advantages and risks. At the consumer-facing application layer, personalization and user experience are the key advantages, but consumer mistrust and misguidance are the corresponding risks, manifesting as information overload and gaps between virtual presentations and real-world products. At the design model layer, intelligent design support and sustainable green design are the advantages, while ethical challenges and socio-cultural risks emerge as design bias, designers’ over-reliance on technology, and erosion of creative subjectivity. At the attribute layer—the data and system foundation—data-driven market analysis and supply chain optimization deliver value, but massive data capture raises privacy and security risks, and the concentration of computing power and data triggers concerns about technological monopoly. The authors emphasize that this dual nature is structural: each layer’s efficiency gains arrive bundled with its own characteristic hazards.

The identified risks are not hypothetical in the reviewed literature. Studies document models that misinterpret a corset as a wide belt based on surface pixels, generate visually plausible garments that omit the darts, plackets, and buttonholes required for physical wearability, and fail to respond consistently to precise instructions such as narrowing a neckline by two centimeters. In education, research cited in the review suggests that while AI improves efficiency, it may weaken students’ creative autonomy, shifting design decisions from problem-driven exploration to result-oriented selection. On the legal side, reviewed work points to a significant lag in laws defining the boundary between human and AI creation and proposes mechanisms such as AI rights registration to address intellectual property questions. Looking forward, the authors argue that isolated optimization of any single layer is insufficient; they propose advancing the three-layer architecture itself and building multi-agent collaborative systems—in which, for example, separate creative director, pattern maker, and market analyst agents negotiate across layers, translating a consumer’s vague request for a breathable, well-draped fabric into precise material parameters and simulation-ready geometry. They also acknowledge their review’s limitations: non-English and non-indexed publications may have been excluded, the sample size may limit generalizability, and qualitative coding is susceptible to researcher subjectivity.

Subject of Research: A systematic literature review of artificial intelligence-generated content applications, technological evolution, advantages, and risks in the fashion industry

Article Title: How AIGC empowers the development of fashion industry: a systematic literature review and future research directions

Article References: Lin, R., Yu, Y., Xia, F., & Guan, M. (2026). How AIGC empowers the development of fashion industry: a systematic literature review and future research directions. Fashion and Textiles, 13(1), Article 32. https://doi.org/10.1186/s40691-026-00479-6

Image Credits: AI Generated

DOI: 10.1186/s40691-026-00479-6

Keywords: AIGC, fashion industry, systematic literature review, diffusion models, generative adversarial networks, large language models, virtual try-on, fashion design, trend forecasting, AI ethics, sustainable fashion, fashion education