A high-performance, GPU-accelerated Virtual Try-On system written in C and Java, leveraging DirectML for real-time human body mapping and garment overlay.
"Bridging computational geometry with deep perceptual understanding for fashion-tech experiences."
If you'd like to contribute or see code examples, contact shaktibiplabDev
This project implements a multi-lingual virtual try-on engine capable of rendering and simulating apparel on human avatars using high-precision body landmark detection. The core pipeline involves:
- Real-time pose estimation
- 3D body mesh alignment
- Physically-based cloth simulation
- DirectML-enhanced parallel rendering
- 🧠 Deep Joint Mapping: Over 60+ anatomical landmark points tracked and interpolated per frame
- ⚙️ Hardware-accelerated: Optimized for AMD GPUs via DirectML (Direct Machine Learning)
- 🧵 Cloth Deformation Model: Mass-spring physics + non-rigid ICP alignment for dynamic garment fitting
- 🧍♂️ Pose-space Deformation (PSD) for context-aware fitment
- 📦 Native C engine for high-performance threading, with Java bindings for UI and logic integration
| GPU VRAM | Approx FPS | Notes |
|---|---|---|
| 4 GB | ~10 FPS | Minimum functional performance |
| 16 GB | ~45 FPS | Recommended for smooth experience |
- Pose Estimation: Optimized using forward kinematics with inverse Jacobian projection
- 3D Alignment: Based on Procrustes analysis for aligning the user silhouette with virtual mesh
- Simulation: Uses Euler integration with Lagrangian multipliers for collision constraints
- Rendering: Real-time rasterization with vertex skinning and morph target blending
- ✅ OS: Windows 10/11
- ✅ GPU: AMD GPU with at least 4 GB VRAM (16 GB recommended)
- ✅ RAM: 8 GB minimum, 16+ GB recommended
- ✅ CMake + GCC (for C components)
- ✅ Java 11+