Jewelry Relighting: Relighting Jewelry via Fine-Tuned Diffusion Models

Yale University
TATA

Examples generated by our method.

Abstract

This project implements a novel jewelry relighting system using a dual-UNet architecture with ControlNet integration. The solution enables photorealistic relighting of jewelry images under various HDR environments while maintaining fine details through edge preservation.

Key Features

Dual-UNet Architecture: Two UNet networks working together for enhanced detail preservation.
ControlNet Integration: Canny edge ControlNet maintains precise jewelry contours during relighting.
HDR Processing: Utilizes logarithmically normalized HDR maps for realistic lighting transfer.
Efficient Fine-tuning: Adapted from MimicBrush framework for resource-efficient training.

Dataset

3D Jewelry Models: 175 objects (variations in silver, gold, and gemstone colors)
Initial lighting setup: front white light for standard jewelry photography.


Viewpoints: 12 camera angles per model