CMU & Meta · Research Paper

SynCD: Multi-Image Synthetic Data for Text-to-Image Customization

SynCD is a data generation pipeline that creates multi-image synthetic datasets of objects in different contexts to train encoder-based models for customized text-to-image generation. The approach uses 3D Objaverse assets, masked shared attention mechanisms, and enforces consistent object identity across generated images, enabling high-quality customization from single or multiple reference images.