Xiaoyu (Nicholas) Wu · 吴晓宇
Hi, I'm Xiaoyu Wu
PhD student · Carnegie Mellon University
I work on generative models, especially long-horizon video generation, representation learning, and reasoning.
News
Scroll for earlier updates
- new New preprints: Compress to Remember and Scaling Video Generation for Reasoning, on compact video memory and the compute cost of reasoning.
- new Also new: Representation by Design in Generation and Visualizing Distribution Coverage, studying what generative models learn and which outputs they cover.
- Started my PhD in the School of Computer Science at Carnegie Mellon University.
- Released DSR, our work on taming outlier tokens in diffusion transformers.
- Selective Jacobi Decoding, our work on accelerating autoregressive normalizing flows, appears in TMLR 2026 with Featured Certification.
- Our work on the corruption stage in few-shot diffusion fine-tuning was accepted to KDD 2026.
- Our work on stealthy backdoor attacks in federated learning appears in ICCV 2025.
- Unlearned but Not Forgotten, our work on extracting removed data after exact unlearning, was accepted to NeurIPS 2025.
- Began PhD studies at Rice University, working with Chen Wei on generative modeling.
- FineXtract, our training-data extraction work for personalized diffusion models, was accepted to ICML 2025.
- Our team won all four tracks of the MIDST Challenge at SaTML 2025. Our membership-inference methods also appear at the TPDP 2025 workshop.
- 2025 Targeted Attack Improves Protection against Unauthorized Diffusion Customization appears at ICLR 2025 as a Spotlight paper.
- Released the FineXtract preprint, showing how model guidance can recover training images from personalized diffusion models.
- Released our corruption-stage preprint, identifying quality degradation in few-shot diffusion fine-tuning and mitigating it with Bayesian neural networks.
- CGI-DM, our work on copyright authentication for diffusion models, was accepted to CVPR 2024. Code is available in Revelio.
- Toward Effective Protection against Diffusion-Based Mimicry through Score Distillation was accepted to ICLR 2024.
- Released Mist-v2, providing enhanced protection against LoRA-based imitation.
- Adversarial Example Does Good, our work on preventing painting imitation with AdvDM, is an Oral paper at ICML 2023.
- Open-sourced Mist, our artwork-protection tool. Read the paper on adversarial examples for diffusion models.
Selected work
Recent work first · Scroll or swipe to explore
Preprint · 2026
Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation
PACC learns compact memory from a frozen video generator, retaining earlier events and improving consistency in long video generation.
Preprint · 2026
Scaling Video Generation for Reasoning: At What Cost?
A controlled Rubik's Cube benchmark reveals the compute cost of hidden-state reasoning and why lower generation loss does not guarantee better reasoning.
Preprint · 2026
Representation by Design in Generation: Cross-View Class-Token Alignment in Diffusion Transformers
Cross-view class-token alignment improves ImageNet linear-probe accuracy by 9.4 percentage points while preserving comparable generation quality.
Preprint · 2026
Visualizing Distribution Coverage in Generative Diffusion Models
Pass@k reveals how distribution-matching distillation trades broad coverage of valid outputs for stronger single-sample quality in image and video generation.
Preprint · 2026
Taming Outlier Tokens in Diffusion Transformers
Dual-Stage Registers reduce harmful outlier tokens in both vision encoders and diffusion transformers, improving generation quality across ImageNet and text-to-image settings.
ICML · 2025
Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models
FineXtract uses the distribution shift between pretrained and personalized diffusion models to recover about 20% of fine-tuning data from real-world checkpoints.
ICML Oral · 2023
Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples
AdvDM uses adversarial examples as protective watermarks that prevent diffusion models from learning an artist's style from unauthorized images.
About
I'm a PhD student at Carnegie Mellon University, where I work with Prof. Niloofar Mireshghallah, Prof. Zhiwei Steven Wu, and Prof. Andrew Ilyas. My research focuses on long-horizon generation, representation learning, and reasoning for generative models. Previously, I pursued PhD studies at Rice University (September 2025–June 2026), working with Prof. Chen Wei.
I founded Mist, an open-source tool for protecting artwork from unauthorized diffusion-model imitation. Our team also won all four tracks of the MIDST Challenge at SaTML 2025.