I am a final-year Ph.D. candidate at MMLab, CUHK, advised by Hongsheng Li and Xiaogang Wang, and currently a Research Intern at ByteDance Seed working with Haoqi Fan. My research focuses on generative models, unified multimodal models, world models, and agentic systems.

Previously, I interned at Reve (with Han Zhang) and Google DeepMind (with Long Zhao), and collaborated with Avolution AI (acquired by MiniMax; with Zhaoyang Huang) and Tencent AI Lab. Across these roles, I worked on generative and multimodal modeling, including Data-Space Iteration, Representation Forcing, Omni, ConsistencySolver, PromptRL, Rectified Diffusion, Diffusion-NPO, Phased Consistency Model (adopted in Qwen-Image-2512 and FastVideo), and AnimateLCM (1M+ Hugging Face downloads), alongside earlier open-source continual learning (PyCIL, 1k+ GitHub stars).

I enjoy unpacking the math and systems behind modern models from first principles—writing deep-dive notes on distributional & trajectory distillation, RL / OPD / VSD density ratios, and long-context attention & KV-cache kernels. I want to stay curious, chase what excites me, and live wholeheartedly.

Papers

Google Scholar ↗

2026 & Recent Preprints

2026

Few-Step Generation via Data-Space Iteration

Shanchuan Lin, Yansong Peng, Fu-Yun Wang, Haoqi Fan

ByteDance Seed

Few-step distillation without a timestep schedule: the generator refines its own image prediction instead of re-noising it.

arXiv Blog

NeurIPS 2026

Representation Forcing for Bottleneck-Free Unified Multimodal Models

Yuqing Wang et al. · Fu-Yun Wang (co-author)

HKU · ByteDance Seed · CUHK · Nanjing University · Tsinghua University

Predicts visual representations before pixels, removing the frozen-VAE bottleneck from unified multimodal models.

Project arXiv

2026

mni: Context Unrolling in mni Models

ByteDance Seed

A unified multimodal model that reasons via context unrolling; core contributor to World Navigation data processing and training validation.

Project arXiv

CVPR 2026 Highlight NVIDIA Tech Talk Tencent Qingyun Travel Grant

Image Diffusion Preview with Consistency Solver

Fu-Yun Wang, Hao Zhou, Liangzhe Yuan, Sanghyun Woo, Boqing Gong, Bohyung Han, Ming-Hsuan Yang, Han Zhang, Yukun Zhu, Ting Liu, Long Zhao

Google DeepMind · CUHK · Seoul National University

Fast, faithful low-step previews that reduce interaction time by nearly 50% while preserving final-sample consistency.

Project arXiv GitHub

ICML 2026 JQ Investments ICML Travel Award

PromptRL: Prompt Matters in RL for Flow-Based Image Generation

Fu-Yun Wang, Han Zhang, Michaël Gharbi, Hongsheng Li, Taesung Park

Reve · Meta Superintelligence Labs · CUHK

2× sample efficiency through joint LM/flow-model RL, plus adaptive test-time prompt refinement.

Project arXiv GitHub

Posts

Refine, Don’t Re-roll: Few-Step Generation by Iterating in Data Space

Our work on few-step distillation without a timestep schedule: the generator refines its own image prediction instead of re-noising it.

LLM 都会画画了，还需要训练视觉生成模型吗？

从一只小猫开始讲起。

So, What is Jev?

Is it really something new?

Mixture of Experts: How Routing Becomes GPU Computation Colab notebook

Why fewer active parameters don’t automatically mean faster models.

RL, OPD, VSD, and GANs: One Density Ratio, Different Updates

How a log-density ratio connects policy gradients, Gaussian mean MSE, diffusion scores, and discriminator feedback—and where the equivalences end.

Scaling Long-Context Attention I: Exact Blocks, Rings, and Sparsity Colab notebook

A unified view of block attention: FlashAttention, Ring Attention, and sparse attention share the same blockwise computation but obey different contracts.

Scaling Long-Context Attention II: From MHA to MLA Colab notebook

How MHA, GQA, MQA, split-KV, Paged Attention, and MLA reshape the KV cache and its decode kernels.

Scaling Long-Context Attention III: Recurrent and Learned Memory Colab notebook

How Linear Attention, chunkwise training, DeltaNet, TTT, and LaCT replace a growing token history with recurrent and learned memory.

Distributional Distillation I: From a Marginal KL to Two Scores

A first-principles derivation of distributional distillation, from marginal KL objectives to two-score and f-divergence gradients.

Distributional Distillation II: What the Methods Actually Estimate

A method-by-method map of what SDS, VSD, Diff-Instruct, DMD, DMD2, and Phased DMD actually estimate.

Distributional Distillation III: From Scores to Particle Flows

From score differences to particle motion, connecting DMD, Drifting Models, and the Sinkhorn-based W-Flow through Wasserstein gradient-flow geometry.

Theoretical Analysis of ConsistencySolver (CVPR 2026 Highlight)

A numerical-analysis view of ConsistencySolver, deriving its learnable update rule from PF-ODEs and linear multistep methods.

Theoretical Analysis of Rectified Diffusion (ICLR 2025)

An optimal-transport analysis extending rectification, path straightening, and O(1/K) convergence from linear flows to general diffusion schedules.

On the Equivalence of Consistency Models and MeanFlow

A continuous-time derivation showing that consistency training reduces to MeanFlow, with an extension from one-step CM to multi-step CTM.

Research

Generated with vibe-manim

An interactive map of the research directions below. Click a node to expand or collapse it; click a paper title to open it.

Interactive research tree

Selected Publications by Theme

Diffusion models: acceleration & reinforcement learning

PromptRL: Prompt Matters in RL for Flow-Based Image Generation

Fu-Yun Wang, Han Zhang, Michaël Gharbi, Hongsheng Li, Taesung Park

Reve · Meta Superintelligence Labs · CUHK

ICML 2026 JQ Investments ICML Travel Award

Jointly trains language models and flow-matching models in a unified RL loop, achieving 2× sample efficiency and adaptive test-time prompt refinement.

Project arXiv GitHub

Image Diffusion Preview with Consistency Solver

Fu-Yun Wang, Hao Zhou, Liangzhe Yuan, Sanghyun Woo, Boqing Gong, Bohyung Han, Ming-Hsuan Yang, Han Zhang, Yukun Zhu, Ting Liu, Long Zhao

Google DeepMind · CUHK · Seoul National University

CVPR 2026 Highlight Tencent Qingyun Travel Grant

A learnable high-order ODE solver for fast, faithful diffusion previews and consistent full-step refinement, reducing interaction time by nearly 50%.

Project arXiv GitHub Theory blog

Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow

Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, Hongsheng Li

ICLR 2025

In-depth theoretical analysis and empirical validation of flow matching, rectified flow, and the rectification operation. ZHIHU blog garnered 10k+ views and ~400 likes. Detailed theory note.

arXiv GitHub Poster Theory blog

Phased Consistency Model

Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, Xiaogang Wang, Hongsheng Li

NeurIPS 2024

Validate the feasibility of one-step video generation; Adopted by Qwen-Image-2512 for first-stage initialization; Adopted by Early version of FastVideo, accelerating HunyuanVideo and WAN.

Project GitHub Paper Poster