Before AI Styling Becomes Mainstream

A retrospective on GAN clothing generation, human visual judgment, and why I've gradually come to believe that generative systems should help people make better choices, not just produce more images.

^{Figure 1: From Clothing Generation to Human-Centered Recommendations. Redrawn in 2026 based on a personal concept proposal I developed in 2019.}

In 2019, I participated in StylingAI, a company project exploring the application of generative artificial intelligence in clothing design. At the time, the visual novelty of AI-generated clothing images was easily noticeable; the real challenge was what these images could actually do for a specific person, a designer, or a fashion company.

This question has stayed with me ever since. Looking back now, I realize that this project gave me more than just a new production tool. This changed the way I thought about problems: I gradually stopped viewing generation itself as the end goal and began to focus on the relationship between visual preferences, body and silhouette, usage context, human judgment, and recommendations.

I wasn't the developer of the underlying model, nor could I claim to understand the specific details of the model architecture. The technical team was using Generative Adversarial Networks (GANs); my contributions mainly came from clothing design execution and cross-team collaboration.

My work included assisting the algorithm team with image annotation, judging the visual credibility of the generated results, managing and reviewing image output, coordinating subsequent modifications and fixes, and helping to connect the technical process with design and business needs. When the generated images were incomplete or the visual effects were unstable, designers still needed to understand, filter, and correct the results, sometimes even requiring a complete rebuild.

This distinction is crucial. I didn't develop the GAN model; my work was at the point where the model output met clothing knowledge and human perception. Because of this position, I saw both the potential of generative technology and its limitations firsthand.

A seemingly plausible image doesn't necessarily equate to useful clothing.

An AI-generated fashion image might seem convincing at first glance, but closer inspection reveals numerous problems: inconsistent proportions, flawed garment structure, and details that are blurred, repetitive, or conflicting. A silhouette might look beautiful, but it's not necessarily wearable, producible, or suitable for a particular body type or occasion.

Therefore, generating an image doesn't mean the work is finished. Human judgment is still required: does the result express a coherent design intent? Does it belong to the required product category? Are the colors and forms harmonious? And can it be further developed into a truly useful design?

By 2020, this exploration extended further to AI and 3D visual presentation. Images at that stage combined GAN-generated materials, 3D production, and post-production adjustments by graphic designers. This process more clearly demonstrated the boundaries between different tasks: machine generation can accelerate visual exploration, but the quality and meaning of the final result still heavily rely on human evaluation and refinement.

During the project, I began to develop my own conceptual framework for intelligent recommendations. I no longer focused solely on the clothing to be generated but placed the user at the center of the system. Around this individual, the system needs to understand a multitude of information: visual preferences, style and patterns, body shape, occasion of use, brand affinity, purchase intent, and broader trend signals.

This idea was still very early and incomplete at the time. It wasn't a finished product specification, much less a validated recommendation model. However, it marked a significant shift in my thinking: if a system only generates a visually appealing image of clothing, it doesn't truly know whether that clothing can genuinely help a specific person. Recommendation requires a different kind of intelligence—connecting visual data with an individual's context, while preserving space for human interpretation and judgment.

In other words, the core issue is no longer how to generate more clothing, but how to relate the generated results to a specific person.

At the time, my team of graphic designers and I understood the refinement of the generated images as a practical step necessary to improve imperfect AI output. Now, I believe its significance is more fundamental. Human judgment not only cleans up the mess left by the model but also provides standards that the model itself doesn't possess independently: coherence, appropriateness, identity expression, confidence, cultural connotations, and sensitivity to context.

These standards are difficult to reduce to mere visual similarity. Two garments may share similar colors, categories, or silhouettes, yet convey entirely different social meanings. A technically sound recommendation may still feel off to the recipient. Therefore, trust cannot be imposed only after the visual interface is complete. It depends on how the system understands the user, interprets the basis for the recommendation, and allows users to disagree and adjust the outcome, retaining their autonomy.

I am currently focusing on AI-assisted visual recommendations for women, particularly the relationship between personalization, visual identity, self-presentation, trust, and agency. The questions are more clearly defined today, but their origins can be traced back to that early experience.

What I am more concerned with now are systems that don't simply label a person or impose a fixed style on the user. A useful recommendation system should help users understand different choices, know why a suggestion appears, and retain control over how they present themselves in different contexts. Style is not just a set of visual attributes; it is also a way for individuals to constantly negotiate between identity, desires, social circumstances, and changes over time.

This perspective also influences my current research prototype, AIFFD. AIFFD explores how to organize body structure, color, personal preferences, and situational needs into a more transparent personal fashion profile. Its goal is not to claim that algorithms can define a person, but to explore how AI can support self-understanding and decision-making without diminishing human agency.

This article is a personal retrospective, not the official technical history of StylingAI. I describe what I have seen, participated in, and learned in my work.

The framework presented in Figure 1 is a redrawing I did in 2026 based on a personal concept proposal that had been developed at that time. It is not the company's original product interface, nor is it a copy of the original project materials. The aim is to make the development trajectory of my personal thinking clearly visible while respecting the contributions and rights of other participants.

Since 2019, generative fashion has changed significantly, but the question I took from that project remains important: For whom is this generation ultimately for?

If the answer is simply an image, then success can be measured by novelty, speed, or visual credibility; if the answer is a person, then the system must also deal with relevance, interpretation, context, trust, and choice. This is a much more difficult problem, but one that I believe is truly worth continuing to study.