Sitemap
A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.
Pages
Xiaoyu (Nicholas) Wu
Xiaoyu Wu is a PhD student at Carnegie Mellon University working on generative models, long-horizon generation, representation learning, and reasoning.
Guidance is becoming a cheat code for LLMs (with Loooop)
Published:
Guidance can dramatically improve generation. Looped models offer a way to reduce the normalization gap between token-level guidance and the guided joint distribution over a full sequence.
Posts
Future Blog Post
Published:
This post will show up by default. To disable scheduling of future posts, edit config.yml and set future: false.
Guidance is becoming a cheat code for LLMs (with Loooop)
Published:
An earlier loop can help guide the final prediction. We explain the decoding rule, the normalization gap between token and sequence guidance, and why close loop logits can narrow that gap.
Compress to Remember
Published:
Prediction-Aligned Context Compaction learns what a frozen video generator should remember. Explore the training method, compact-memory inference, and matched video comparisons.
EMA: A Quiet Hyperparameter That Moves Diffusion Leaderboards Permalink
Published:
EMA decay can change diffusion-model rankings and the balance between sample fidelity and coverage. We examine its effect across pixel-, latent-, and representation-space models.
Blog Post number 4
Published:
This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.
Blog Post number 3
Published:
This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.
Blog Post number 2
Published:
This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.
Blog Post number 1
Published:
This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.
publications
Mist: Towards Improved Adversarial Examples for Diffusion Models
Published in arXiv preprint, 2023
Mist improves the transferability and robustness of adversarial watermarks designed to protect artwork from unauthorized diffusion-model imitation.
Recommended citation: Liang, C.* and Wu, X.* (2023). "Mist: Towards Improved Adversarial Examples for Diffusion Models." arXiv:2305.12683.
Download Paper
Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples
Published in ICML 2023 (Oral), 2023
AdvDM uses adversarial examples as protective watermarks that prevent diffusion models from learning an artist’s style from unauthorized images.
Recommended citation: Liang, C.*, Wu, X.*, Hua, Y., et al. (2023). "Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples." ICML 2023. Oral presentation.
Download Paper
Toward Effective Protection against Diffusion-Based Mimicry through Score Distillation
Published in ICLR 2024, 2024
Score distillation makes protective image perturbations faster and more memory-efficient while strengthening defenses against unauthorized diffusion-based mimicry.
Recommended citation: Xue, H., Liang, C.*, Wu, X.*, and Chen, Y. (2024). "Toward Effective Protection against Diffusion-Based Mimicry through Score Distillation." ICLR 2024.
Download Paper
CGI-DM: Digital Copyright Authentication for Diffusion Models via Contrasting Gradient Inversion
Published in CVPR 2024, 2024
CGI-DM visualizes conceptual differences between pretrained and fine-tuned diffusion models to provide evidence for copyright authentication.
Recommended citation: Wu, X., Hua, Y., Liang, C., et al. (2024). "CGI-DM: Digital Copyright Authentication for Diffusion Models via Contrasting Gradient Inversion." CVPR 2024.
Download Paper
Targeted Attack Improves Protection against Unauthorized Diffusion Customization
Published in ICLR 2025 (Spotlight), 2025
Carefully selected targeted attacks provide stronger protection against unauthorized diffusion customization than untargeted adversarial watermarks.
Recommended citation: Zheng, B.*, Liang, C.*, and Wu, X. (2025). "Targeted Attack Improves Protection against Unauthorized Diffusion Customization." ICLR 2025. Spotlight presentation.
Download Paper
Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models
Published in ICML 2025, 2025
FineXtract uses the distribution shift between pretrained and personalized diffusion models to recover about 20% of fine-tuning data from real-world checkpoints.
Recommended citation: Wu, X., Zhang, J., and Wu, S. (2025). "Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models." ICML 2025.
Download Paper
Winning the MIDST Challenge: New Membership Inference Attacks on Diffusion Models for Tabular Data Synthesis
Published in TPDP 2025, 2025
A lightweight learned membership inference attack reveals privacy leakage in diffusion-based tabular synthesis and won all four tracks of the MIDST Challenge.
Recommended citation: Wu, X., Pang, Y., Liu, T., and Wu, Z. S. (2025). "Winning the MIDST Challenge: New Membership Inference Attacks on Diffusion Models for Tabular Data Synthesis." TPDP 2025.
Download Paper
Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM
Published in NeurIPS 2025, 2025
Exact unlearning can still leak removed examples when pre- and post-unlearning models are exposed; our guided attack doubles extraction success in some settings.
Recommended citation: Wu, X., Pang, Y., Liu, T., and Wu, Z. S. (2025). "Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM." NeurIPS 2025.
Download Paper
Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment
Published in ICCV 2025, 2025
Adaptive layer-wise gradient alignment exposes weaknesses in federated-learning defenses by preserving backdoor effectiveness while making malicious updates harder to detect.
Recommended citation: Yang, Q., Yan, P., Wu, X., et al. (2025). "Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment." ICCV 2025.
Download Paper
Exploring Diffusion Models’ Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks
Published in KDD 2026, 2026
We identify a corruption stage during few-shot diffusion fine-tuning and use Bayesian neural networks to improve fidelity, quality, and diversity without extra inference cost.
Recommended citation: Wu, X.*, Zhang, J.*, Hua, Y., et al. (2026). "Exploring Diffusion Models' Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks." KDD 2026.
Download Paper
Accelerating Inference of Discrete Autoregressive Normalizing Flows by Selective Jacobi Decoding
Published in TMLR 2026 (Featured Certification), 2026
Selective Jacobi decoding speeds up autoregressive normalizing-flow inference by up to 4.7 times while preserving generation quality, without additional training.
Recommended citation: Zhang, J., Lu, J., Wu, X., Wang, Z., and Zhang, R. (2026). "Accelerating Inference of Discrete Autoregressive Normalizing Flows by Selective Jacobi Decoding." Transactions on Machine Learning Research. Featured Certification.
Download Paper
Taming Outlier Tokens in Diffusion Transformers
Published in arXiv preprint, 2026
Dual-Stage Registers reduce harmful outlier tokens in both vision encoders and diffusion transformers, improving generation quality across ImageNet and text-to-image settings.
Recommended citation: Wu, X.*, Wang, Y.*, Fu, T.-J., et al. (2026). "Taming Outlier Tokens in Diffusion Transformers." arXiv:2605.05206.
Download Paper
Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation
Published in arXiv preprint, 2026
PACC learns compact memory from a frozen video generator, retaining earlier events and improving consistency in long video generation.
Recommended citation: Wu, X.*, Guo, W.*, Wang, Y.*, Feng, X., Kavraki, L. E., and Wu, Z. S. (2026). "Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation." arXiv:2609.36364.
Download Paper
Representation by Design in Generation: Cross-View Class-Token Alignment in Diffusion Transformers
Published in arXiv preprint, 2026
Cross-view class-token alignment improves ImageNet linear-probe accuracy by 9.4 percentage points while preserving comparable generation quality.
Recommended citation: Wu, X., Wang, Y., and Wei, C. (2026). "Representation by Design in Generation: Cross-View Class-Token Alignment in Diffusion Transformers." arXiv:2609.36348.
Download Paper
Scaling Video Generation for Reasoning: At What Cost?
Published in arXiv preprint, 2026
A controlled Rubik’s Cube benchmark reveals the compute cost of hidden-state reasoning and why lower generation loss does not guarantee better reasoning.
Recommended citation: Guo, W., Wu, X., Wang, Y., Mireshghallah, N., and Kavraki, L. E. (2026). "Scaling Video Generation for Reasoning: At What Cost?" arXiv:2609.36599.
Download Paper
Visualizing Distribution Coverage in Generative Diffusion Models
Published in arXiv preprint, 2026
Pass@k reveals how distribution-matching distillation trades broad coverage of valid outputs for stronger single-sample quality in image and video generation.
Recommended citation: Wang, Y., Wu, X., Fu, T.-J., Chen, C., Chen, L.-C., Gan, Z., and Wei, C. (2026). "Visualizing Distribution Coverage in Generative Diffusion Models." arXiv:2609.38853.
Download Paper