IIT-R unveils India’s first text to fashion sketch AI system

Roorkee: Researchers at the Indian Institute of Technology Roorkee (IIT-R) have developed a new technology that enables AI systems to generate high-fidelity fashion sketches directly from textual descriptions, a first-of-its-kind initiative in India. Until now, most AI applications in the fashion sector have relied on photographs and catalogue images. While effective for online retail, product search and recommendations, such systems fail to reflect how fashion designers actually conceptualise ideas.

Designers primarily work through sketches and industry-specific terminology rather than finished product images, making it difficult for existing AI tools to interpret complex design descriptions beyond basic keywords.


To address this gap, the researchers developed FLORA (Fashion Language Outfit Representation for Apparel Generation), the country's first comprehensive fashion sketch dataset created specifically for text-to-outfit generation. The dataset comprises 4,330 curated pairs of professional fashion sketches and corresponding textual descriptions written using terminology commonly used by designers. These descriptions capture fine-grained details such as silhouette, fabric, texture, construction and overall style.

Unlike conventional photo-based datasets, FLORA mirrors what designers typically draw in their sketchbooks. As a result, AI models trained on it can interpret not only simple prompts such as "a blue dress" but also complex descriptions like "a cobalt blue evening gown with a softly draped cowl neckline and a ruched bodice," enabling more precise and expressive design generation at the conceptual stage.

However, the researchers noted that a dataset alone was insufficient to achieve accurate results. Existing parameter-efficient fine-tuning techniques such as LoRA, LoKR, DoRA and LoHA rely on linear adapters and often struggle to capture complex semantic relationships in design language.

"To overcome this limitation, we introduced NeRA (Nonlinear low-rank Expressive Representation Adapter), a novel adapter architecture based on Kolmogorov-Arnold Networks," said professor Sparsh Mittal, principal investigator of the study from IIT Roorkee's Mehta Family School of Data Science and Artificial Intelligence. "NeRA uses learnable nonlinear transformations that help AI models better capture subtle stylistic intent, achieve stronger semantic alignment and converge faster during training."

The research was a collaborative effort between Gayatri Deshmukh (independent researcher), Somsubhra De (IIT Madras), Chirag Sehgal (Delhi Technological University) and Jishu Sen Gupta (IIT BHU).

Researchers said the framework has the potential to accelerate design workflows, enable greater personalisation and support sustainable fashion practices by reducing trial and error in the early stages of design.