Multimodal researchers can test alignment and retrieval when collecting matched image-text data is difficult. Core capabilities include Unpaired representation alignment, Image-text experiments, Wasserstein Procrustes mapping, and Orthogonal transformation.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Coarse geometric initialization, Zero-shot classification evaluation, Cross-modal retrieval, and Public reference implementation.
