teaches machines to denoise reality.
Final-year Statistics PhD at UT Austin, advised by Prof. Mingyuan Zhou. I build diffusion models and multimodal generative systems — and the algorithms that keep them fast, safe, and trustworthy.
Recent updates
What I work on
I'm interested in the theory and practice of generative modeling — particularly how iterative denoising processes can be generalized beyond the Gaussian playbook and pushed toward faster, safer, and more controllable systems. My recent work spans four threads:
A unified, Bregman-divergence view of iterative corruption / recovery — leading to non-Gaussian variants like Beta Diffusion (NeurIPS '23) and Learning to Jump (ICML '23) for sparse, non-negative, heavy-tailed data.
Score Forgetting Distillation (ICLR '25): a swift, data-free way to forget unsafe classes or concepts (incl. specific celebrities and NSFW content) while preserving generation quality — and getting up to 1000× sampling speedup for free.
A dense-reward perspective on aligning text-to-image diffusion with human preference — turning the trajectory itself into the optimization signal (ICML '24).
iPromptDiff: an SD-based architecture that decouples content from task and routes visual perception through text embeddings — strong in-domain and OOD performance even when text prompts are missing.
Papers
A teacher–student distillation that rapidly removes target classes or concepts from diffusion models without accessing real data, while preserving overall generative quality.
Trajectory-level dense reward signals for aligning text-to-image diffusion to human preference, outperforming sparse-reward baselines.
A diffusion model defined on the simplex via Beta distributions — well-suited for bounded data such as images and probability vectors.
A binomial / Poisson hierarchical VAE that handles sparsity, skewness, heavy tails, and heterogeneity — natural for count-like and non-negative data.
An information-theoretic surrogate for DkNN classification, plus matching attack and defense algorithms with state-of-the-art adversarial robustness.
An immune-system-inspired adversarial framework that defends against unseen attacks by mimicking B-cell affinity maturation.
Where I've worked
Built a full data-curation and training pipeline for diffusion-based video super-resolution; designed temporally-consistent simulators of real-world video degradation; developed a VSR diffusion model on CogVideoX 1.5.
Designed an automatic short-form video localization pipeline (object detection, OCR, inpainting, segmentation), and a landscape-to-portrait conversion workflow with KMeans + Gaussian-process smoothing for stable subject tracking.
Studied visual in-context learning for diffusion models; proposed iPromptDiff, an SD-based architecture that decouples content from task and pushes visual perception into text embeddings — beating baselines in-domain and OOD.
Built and open-sourced a PyTorch codebase for DDPM, DDIM, and classifier-free guidance (230★+ on GitHub), reproducing published diffusion results and exploring non-Gaussian generalizations.
Co-developed adversarial soft k-NN (ASK) and the immune-inspired RAILS framework for adversarial robustness on deep classifiers.
Education
Advised by Prof. Mingyuan Zhou, IROM, McCombs School of Business.
Awards & service
Fellowships & awards
- University Graduate Continuing Fellowship 2025
- McCombs Dean's Fellowship 2022 – 2024
- NeurIPS Scholar Award 2023
- UT Professional Development Award 2023
- Fudan Excellent Freshman Scholarship — Top 1% 2015
Service
- Reviewer · ICLR '24 – '26
- Reviewer · NeurIPS '23 – '25
- Reviewer · ICML '23 – '25
- Reviewer · AISTATS '21, '26
- Teaching Assistant · UT Austin '21 – '22, '24
Toolbox
- Python
- PyTorch
- JAX
- TensorFlow
- NumPy / SciPy
- scikit-learn
- R
- CUDA
- Diffusion models
- VAEs
- RLHF / DPO
- CogVideoX · SD · Flow Matching
Let's talk.
I'll be on the 2026 academic and industry job market. If you're hiring, collaborating, or just want to argue about score-based vs. flow-based generative models — say hi.