Publication
Hierarchical Cross-modal Representation Alignment for Text-to-image Diffusion Models.
First author · Under review at ICLR 2026
We introduce Hierarchical Intermodal Projection (HIP), a lightweight method for fine-tuning text-to-image diffusion models. HIP aligns phrases with image regions in a shared hyperbolic space while anchoring local representations to their global scene context. It improves compositional generation on SD3.5-M and FLUX.1-dev without adding inference overhead.
MIRROR: Aligning Semantic Relations from Language to Image via Gromov-Wasserstein
ECCV 2026 · Equal contribution
Hong-Han Wang*, Yuntao Wang*, Hu Ding
We introduce MIRROR, a geometric regularization framework that transfers relational priors from language to visual representations in multimodal large language models. A Semi-Inverse Gromov-Wasserstein (SI-GW) formulation yields a closed-form target for visual relational structure, enabling efficient training. MIRROR improves relational consistency while preserving general vision-language performance, without adding parameters or inference overhead.
Finding Wasserstein Ball Center: Efficient Algorithm and The Applications in Fairness
ICML 2025
Yuntao Wang, Yuxuan Li, Qingyuan Yang, Hu Ding
We study the Wasserstein ball center, which minimizes the maximum Wasserstein distance to a collection of distributions. For fixed support, we develop an efficient interior-point algorithm and demonstrate applications to fairness across input distributions.