About Me

MCLR: Improving Conditional Modeling in Visual Generative Models via Inter-Class Likelihood-Ratio Maximization

Xiang Li, Yixuan Jia, Xiao Li, Jeffrey Fessler, Rongrong Wang, Qing Qu

arXiv preprint, 2026

paper

We propose MCLR, a principled alignment objective that maximizes inter-class likelihood-ratios during training, enabling diffusion models to achieve classifier-free guidance-like improvements under standard sampling without inference-time guidance. We further establish a formal equivalence between CFG and alignment-based objectives.

FlowDAS: A Stochastic Interpolant-Based Framework for Data Assimilation

Siyi Chen*, Yixuan Jia*, Qing Qu, He Sun†, Jeffrey Fessler

NeurIPS, 2025

paper / code

We introduce FlowDAS, a generative data assimilation framework that uses stochastic interpolants to learn observation-conditioned state transition dynamics from data, enabling step-by-step state estimation for stochastic dynamical systems without requiring known physical models.

Shorter SPECT Scans Using Self-Supervised Coordinate Learning to Synthesize Skipped Projection Views

Zongyu Li*, Yixuan Jia*†, Xiaojian Xu, Jason Hu, Yuni Dewaraja, Jeffrey Fessler

EJNMMI Physics, 2025

paper / code

We adapt the neural radiance field (NeRF) concept to SPECT imaging, enabling significant reduction in acquisition time (by 2×, 4×, or 8×) via self-supervised coordinate learning to synthesize skipped projection views.

Y90 SPECT Scatter Estimation and Voxel Dosimetry Using a Unified Deep Learning Framework

Yixuan Jia†, Zongyu Li, Azadeh Akhavanallaf, Jeffrey Fessler, Yuni Dewaraja

EJNMMI Physics, 2023

paper / code

We developed a unified three-stage deep learning framework for clinical Y90 SPECT imaging: CNN-based scatter estimation, SPECT reconstruction with scatter correction, and dose-rate map generation.