FitForge AI Fashion Platform
overview
FitForge is an AI-powered fashion platform that uses parametric body modelling to generate a digital avatar from user measurements, then overlays clothing items in augmented reality so users can see how garments fit their specific body shape before purchasing. Built with Team Nexus as a side project alongside Master of IT coursework.
The body modelling pipeline uses SMPL-X (Skinned Multi-Person Linear model with expression parameters), a parametric model that represents body shape and pose as a low-dimensional vector of body shape coefficients and joint rotations. From user-provided measurements (height, chest, waist, hip), the system infers the closest SMPL-X parameters and renders a 3D avatar in a browser-based AR view.
The recommendation layer takes the avatar's body parameters plus user style preferences and queries a product catalog with AI-assisted fit prediction estimating how a specific garment's sizing will translate to the avatar body shape. This addresses the core problem of online fashion: fit uncertainty is the primary driver of returns.
architecture
User Measurements
(height, chest, waist, hip)
│
▼
SMPL-X Parameter Inference
└── body shape vector β ∈ ℝ¹⁰
│
▼
3D Body Mesh Generation
└── 10,475 vertices, 20,908 faces
│
▼
AR Renderer (WebGL)
└── overlay garment mesh on body avatar
│
▼
Fit Prediction Model
└── garment size × body params → fit score
│
▼
AI Recommendation Engine
└── style preference × fit score → ranked catalogSMPL-X represents body shape as a linear combination of principal components learned from a large corpus of 3D body scans. This means any body shape can be approximated as a weighted sum of "shape basis" vectors. The body shape space is compact 10 parameters capture ~95% of human body shape variation making inference from measurements tractable.
technical.decisions
Building a parametric body model from scratch would require a large corpus of 3D body scans and weeks of training. SMPL-X is a published, pre-trained model with a well-documented parameter space. Using it as the body representation lets the project focus on the application layer fit prediction and recommendation rather than body modelling research.
A native iOS/Android app would access ARKit/ARCore for better tracking but requires app store distribution and platform-specific development. A WebGL-based AR view runs in any modern browser, enabling rapid iteration and easy sharing during the prototype phase. The fidelity trade-off is acceptable for proof-of-concept.
Standard size charts (S/M/L/XL) are inconsistent across brands and meaningless for non-standard body proportions. A parametric fit model that takes garment measurements and body parameters as input can predict fit quality more accurately than size lookup, and can generate specific size recommendations per body shape.
outcomes
- SMPL-X avatar generation from user measurements
- WebGL AR overlay for garment visualization in-browser
- Parametric fit prediction body shape aware, not size-chart lookup
- AI recommendation layer combining style preference and fit score