Education
Research Experience
Publications
IMS³: Breaking Distributional Aggregation in Diffusion-Based Dataset Distillation
Chenru Wang*; Yunyi Chen*; Zijun Yang; Joey Tianyi Zhou; Chi Zhang#. (* equal contribution, # corresponding author)
IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026 (Accepted)
Developed Inversion-Matching (IM) to align denoising and inversion trajectories, broadening distributional coverage of distilled samples. Designed Selective Subgroup Sampling (S³) to improve inter-class separability and boundary coverage, achieving state-of-the-art performance among diffusion-based distillation methods.
MGPO: Manifold-Guided Diffusion Alignment for Task-Aware Dataset Distillation
Yunyi Chen*; Chenru Wang*; Xinyi Ye; Zexin Zheng; Chi Zhang#. (* equal contribution, # corresponding author)
Conference on Neural Information Processing Systems 2026 (Under Review)
Proposed Manifold-Guided Policy Optimization (MGPO), formulating diffusion-based dataset distillation as a multi-objective reinforcement learning problem. Combined pixel-space discriminative rewards with latent-space manifold rewards guided by class-wise MST skeletons, improving downstream classification, detection, and segmentation performance.
Skills
Awards & Scholarships
Course Projects
- Developed a geometry-guided pipeline that lifts exocentric frames into 3D and warps them along target head-camera trajectories to provide partially observed egocentric priors
- Fine-tuned a pretrained video-diffusion model with LoRA and autoregressive history conditioning, enabling coherent long-horizon egocentric video generation
- Implemented a bridge-worker framework on Stable Diffusion 1.5, supporting latent-window slicing, spherical prompt interpolation, and multi-prompt panoramic composition
- Evaluated Bridge Correction and SMC/FKC-PoE sampling across 50+ configurations, improving the reliability of seamless transitions between generated regions
- Achieved real-time state sync and visualization in Unity using the Default Visualization Suite: scans, map, local/global costmaps, route plan, and goal positions
- Implemented hybrid obstacle avoidance: ROS2 SLAM/Nav stack + Unity NavMesh, with runtime switching and goal-skip for unreachable targets