Teaching DNNs to design fast fashion

Abhinav Ravi, Arun Patro, Vikram Garg, Anoop Kolar Rajagopal, Aruna Rajan, Rajdeep Hazra BanerjeearXiv 2019 · OpenReview

Abstract

"Fast Fashion" spearheads the biggest disruption in fashion that enabled to engineer resilient supply chains to quickly respond to changing fashion trends. The conventional design process in commercial manufacturing is often fed through "trends" or prevailing modes of dressing around the world that indicate sudden interest in a new form of expression, cyclic patterns, and popular modes of expression for a given time frame. In this work, we propose a fully automated system to explore, detect, and finally synthesize trends in fashion into design elements by designing representative prototypes of apparel given time series signals generated from social media feeds. Our system is envisioned to be the first step in design of Fast Fashion where the production cycle for clothes from design inception to manufacturing is meant to be rapid and responsive to current "trends". It also works to reduce wastage in fashion production by taking in customer feedback on sellability at the time of design generation. We also provide an interface wherein the designers can play with multiple trending styles in fashion and visualize designs as interpolations of elements of these styles. We aim to aid the creative process through generating interesting and inspiring combinations for a designer to mull by running them through her key customers.

Introduction

The "Fast Fashion" industry lives and dies by its ability to respond to rapidly shifting trends. Traditionally, this process is highly manual and subjective: designers browse runways and street style, create mood boards, and produce physical prototypes. This manual approach is slow, prone to human bias, and often leads to significant environmental waste when misjudged trends result in unsold inventory.

Researchers from Myntra (a major Indian fashion e-commerce platform) presented a solution at the KDD 2019 Workshop on AI for Fashion. They proposed an end-to-end pipeline that uses Deep Neural Networks (DNNs) to detect fashion trends from social media and automatically synthesize new apparel designs using Neural Style Transfer.

The Proposed Pipeline

The system aims to replace subjective "gut feeling" with a data-backed system that can quickly generate and validate design prototypes before a single piece of fabric is cut.

The end-to-end pipeline for automated fashion trend detection and design synthesis.

1. Trend Detection

The system crawls Instagram feeds of fashion influencers and uses a pre-trained VGG-19 network to generate image embeddings, which are then clustered. By analyzing these clusters over time, the system identifies three distinct categories of trends: "seasonal trends" (stable popularity), "fads" (sudden spikes), and "increasing trends" (gradual growth).

Classification of fashion trends into seasonal, sudden (fads), and increasing categories based on social media signals.

2. Fashion Generation (Neural Style Transfer)

Once a trend (like a specific pattern or color) is identified, it is treated as a "style" image. A basic apparel silhouette (like a t-shirt or dress) is used as the "content" image. The system employs the Gatys et al. Neural Style Transfer algorithm to merge the two. An L-BFGS optimizer is used to minimize a weighted loss function of content and style.

3. Super Resolution

Because high-resolution style transfer is computationally expensive, the system initially generates low-resolution designs. It then uses a Super-Resolution Convolutional Neural Network (SRCNN) to upscale these images to 2400x2400 pixels (approx. 400 DPI) for professional visualization.

Key Insights and Results

One of the most significant findings was the emergence of "creative" interpretations by the network. For instance, when applying a "fish scale" pattern to a sheath dress, the AI interpreted the scales at the hemline to create a ruffled, "mermaid-like" silhouette, mimicking a structural change without explicit instruction.

The study also highlighted a "Designer-Customer Gap." Professional designers often rejected bold combinations (like neon colors on snake prints) as "overpowering," while customers rated them as "fresh" and "quirky." This suggests AI can help designers step outside their own stylistic biases.

In qualitative validation with 15 designers and 10 customers:

  • 6 out of 15 designs received high scores for both "inspiration" and "manufacturability."
  • 5 designs were seen as highly inspiring but difficult to manufacture.
  • The style transfer network takes about 2 minutes to generate a 600x600 image on an NVIDIA K80 GPU.

Limitations and Future Work

While the AI generates visually stunning 2D images, it does not yet understand 3D garment construction or the physical properties of different fabrics. Currently, human designers are still required to "clean up" the AI's output and add final creative touches before production. Future work involves automating the weighting parameters for content vs. style loss based on specific fashion categories.

Summary updated

Abstract

Fast Fashion" spearheads the biggest disruption in fashion that enabled to engineer resilient supply chains to quickly respond to changing fashion trends. The conventional design process in commercial manufacturing is often fed through "trends" or prevailing modes of dressing around the world that indicate sudden interest in a new form of expression, cyclic patterns, and popular modes of expression for a given time frame. In this work, we propose a fully automated system to explore, detect, and finally synthesize trends in fashion into design elements by designing representative prototypes of apparel given time series signals generated from social media feeds. Our system is envisioned to be the first step in design of Fast Fashion where the production cycle for clothes from design inception to manufacturing is meant to be rapid and responsive to current "trends". It also works to reduce wastage in fashion production by taking in customer feedback on sellability at the time of design generation. We also provide an interface wherein the designers can play with multiple trending styles in fashion and visualize designs as interpolations of elements of these styles. We aim to aid the creative process through generating interesting and inspiring combinations for a designer to mull by running them through her key customers.
BibTeX
@misc{ravi2019teaching,
  author = {Abhinav Ravi and Arun Patro and Vikram Garg and Anoop Kolar Rajagopal and Aruna Rajan and Rajdeep Hazra Banerjee},
  title = {{Teaching DNNs to design fast fashion}},
  year = {2019},
  eprint = {1906.12159},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1906.12159},
}
Figure