BirDress

Exploration of contemporary fashion design through StyleGAN, utilizing curated datasets of birds and dresses for generative modeling

2021

Generative AI DesignFashion TechnologyComputational CreativityMachine Learning
I. CONCEPT

OVERVIEW

BirDress is an exploratory project that uses StyleGAN to draw visual parallels between organic avian forms and contemporary haute couture. It investigates how machine learning can translate structural and textural cues from nature into fashion-adjacent imagery.

MOTIVATION

The project asks how organic life forms might inform design language. While an ecological focus on extinct species was considered, limited imagery shifted the emphasis toward visual experimentation and dataset curation.

APPROACH

Training was performed with curated bird and fashion image sets. To stabilize outputs, datasets were preprocessed for axial alignment so aspect ratios and orientations corresponded before StyleGAN training.

II. DATA

DOMAINS

Reference domains map for BirDress

DATASETS

Bird imagery: Caltech-UCSD Birds 200 (CUB-200-2011) refined to 83 curated images across visually distinctive species.

Fashion imagery: 400 images collected from Iris van Herpen’s haute couture collections for their kinetic, organic qualities.

Curated bird dataset montage

MODEL TRAINING

Early runs with single-domain data under-performed. Preprocessing both datasets to align orientation and crop yielded more coherent hybrid outputs, indicating visual structure and alignment can matter more than dataset size at short training horizons.

III. RESULTS

6 HOURS

9 HOURS

IV. POSTERS
VI. INSIGHTS

TECHNICAL

Visual structure and axial alignment impacted short-horizon training more than raw dataset size.

CONCEPTUAL

Intentional curation enables StyleGAN to generate compelling abstractions that merge fashion and nature imagery.