Fu-Yun Wang (IPA: [fu˧˥ yn˧˥ wɑŋ]) is a third-year Ph.D. candidate at MMLab@CUHK. He works on post-training. Currently vibing with world model and agentic stuff.
Job market 2027. Looking for opportunities (industry) to work on interesting projects. Always happy to chat early.
Recent Work New
Method
ConsistencySolver is a lightweight, learnable high-order ODE solver derived from general linear multistep methods, optimized via Reinforcement Learning (PPO). It adapts integration coefficients to each sampling step, improving both preview fidelity and consistency with the full-step output—without modifying the base diffusion model.
- Diffusion Preview Paradigm: Fast few-step preview → user evaluation → full-step refinement only when satisfied
- Learnable ODE Solver: Adaptive multistep coefficients predicted by a lightweight MLP, trained end-to-end via RL
- Preview Consistency: Preserves the deterministic PF-ODE mapping so preview closely matches the final output
- FID on-par with 47% fewer steps vs. Multistep DPM-Solver; outperforms distillation baselines
Method
PromptRL jointly trains language models and flow-matching models in a unified RL loop for text-to-image generation, using LMs as adaptive prompt refiners.
- Unified RL Training: Joint optimization of both LM and FM in a single loop
- Exploration Collapse Solution: Overcomes insufficient diversity through adaptive prompting
- Test-time Adaptation: Adaptive prompt refinement at inference
Blogs
Detailed derivation of the simplified PF-ODE via variable substitution, background on Linear Multistep Methods, and how ConsistencySolver's learnable update rule unifies DDIM, PNDM, and DPM-Solver-2 as special cases. Companion note for ConsistencySolver (CVPR 2026).
Self-contained mathematical proofs showing that any general diffusion form achieves non-increasing optimal transport costs through rectification — generalizing Rectified Flow beyond the linear interpolation form. Companion note for Rectified Diffusion (ICLR 2025).
Proves that the Consistency Model objective in the continuous-time limit reduces to MeanFlow, generalizes to arbitrary target time $s$, and validates MeanFlow on diffusion policy for robotic control with 1-step generation matching 100-step sampling.
Research Summary
Interactive tree diagram of my research. Click nodes to expand/collapse; click paper titles to visit links.
Research Directions
Internship Experience
ByteDance Seed Current
Research Intern · 2025.10 - Present
Video Generation, Multimodal Models
Mentor: Haoqi Fan
Reve Inc
Research Intern · 2025.6 - 2025.11
Multimodal LMs, Diffusion Models, RL
Supervised by: Dr. Han Zhang
Google DeepMind
Research Intern · 2025.2 - 2025.5
Diffusion Distillation, RL
Supervised by: Dr. Long Zhao, Dr. Ting Liu, Dr. Hao Zhou, Dr. LiangZhe Yuan
Collaborated with Prof. Bohyung Han, Prof. Boqing Gong
Avolution AI Acquired by MiniMax
Research Collaboration · 2023.10 - 2024.10
Video Diffusion, Distillation
Collaborated with: Dr. Zhaoyang Huang, Dr. Xiaoyu Shi, Weikang Bian
Tencent AI Lab
Research Intern · 2022.6 - 2022.12
Class-Incremental Learning
Supervised by: Dr. Liu Liu, Prof. Yatao Bian
Education
The Chinese University of Hong Kong
Ph.D. in Engineering · 2023 - Present
Supervisor: Prof. Hongsheng Li & Prof. Xiaogang Wang
Nanjing University
B.Eng. in AI (Rank 2/88) · 2019 - 2023
Supervisor: Prof. Han-Jia Ye & Prof. Da-Wei Zhou
Selected Publications
Categorized by theme. Full list on Google Scholar.
Diffusion Post-Training: Acceleration & Reinforcement Learning
Generative Vision Applications
Class-Incremental Learning
Talks
Awards & Honor
- 2025 CVPR 2025 Outstanding Reviewer
- 2023 HKPFS (Hong Kong PhD Fellowship Scheme)
- 2023 Outstanding Graduate of Nanjing University
- 2023 Outstanding Undergraduate Thesis of NJU
- 2022 Sensetime Scholarship
- 2022 Huawei Scholarship
- 2021 National Scholarship
Academic Services
Reviewer for:
| Venue | Years |
|---|---|
| TPAMI | — |
| TCSVT / PRL | — |
| CVPR | 2023–2025 |
| NeurIPS | 2023, 2025 |
| ICLR | 2024, 2025 |
| ICML | 2024, 2025 |
| ECCV / ICCV | 2024 / 2025 |
| BMVC / SIGGRAPH Asia | 2024 / 2025 |
