Xiang Wen

Building real-time interactive virtual worlds

Print this page to export a PDF

Summary

Working toward real-time interactive virtual worlds, across generative AI, diffusion models, computer graphics, and 3D vision.
PhD in Artificial Intelligence from Zhejiang University. Nine years of industry research and engineering leadership, moving techniques across game AI, 3D and animation generation, and large-scale video generation.
Focused on turning vague product goals into post-training tasks that can actually be trained, evaluated, and shipped: data construction, supervised fine-tuning, reward modeling, preference optimization, and closed-loop online evaluation.
Led algorithm research and team building for video, 3D, and game AI at NetEase, ByteDance, SkyReels AI, and Tencent, repeatedly building new directions from scratch and taking models to production.
Published at TPAMI, ICML, IEEE TMM, SIGGRAPH Asia, and ACM MM, and serves as a regular reviewer.

Interactive worlds: Consistent world state
Interactive generation
Latency and streaming
Long-horizon controllability
Post-training and alignment: Data construction and mixture
Supervised fine-tuning
Reward modeling
DPO / Flow-DPO / RLHF
Reasoning data and chain-of-thought
Evaluation and online feedback loops
Generative models: Diffusion / DiT
Mixture of experts
Autoregressive generation
Controllable generation
Inference acceleration and distillation
Reinforcement learning: Policy optimization
Hierarchical RL
Policy regularization under dynamics shift
Multi-agent games
3D and animation: 3D generation
Speech-driven facial animation
Stylized avatar creation
Motion generation

Jul 2025 - Present

T12 Senior Researcher · Shenzhen

Lead post-training for anime-domain video generation, organizing data and training around text grounding, motion plausibility, character performance, and style consistency.
Drive a unified video generation architecture supporting text-to-video, image-to-video, multi-reference generation, and video editing within a single model.
Drive inference acceleration and distillation, preserving generation quality while substantially reducing sampling steps.

Jul 2023 - Jul 2025

Business Lead and CTO · Beijing

Built SkyReels from scratch, designing the full creation pipeline from script generation to storyboarding, shot planning, and video generation, through to commercialization.
Designed a multi-agent framework for scripting, storyboarding, and shot planning; built chain-of-thought data and ran supervised fine-tuning plus reinforcement learning on a reasoning model.
Led post-training and preference alignment for video generation: supervised fine-tuning, reward modeling, preference optimization, and a badcase feedback loop.
Led the open-source releases of SkyReels-V1 and SkyReels-A1, and the development of the VolGen 3D generation model, published in TPAMI.
Designed a hierarchical multimodal agent architecture for a game-development agent engine, with a tiered benchmark to guide iteration.

Oct 2020 - Jul 2023

Game AI Technical Lead · Hangzhou

Proposed a GAN-inversion approach to stylized avatar creation, reframing character customization as latent-space search with mixed continuous and discrete parameter optimization. Published at SIGGRAPH Asia and shipped in a live game.
Proposed a speech-driven 3D facial animation method combining a sparsity regularizer with a non-autoregressive backbone, improving lip sync and expression naturalness on long sequences. Published at ACM MM.
Built a semantic-graph-matching framework for full-body motion generation, enabling automatic generation of narrative animation. Invited to present the work at the Game Developers Conference.

Jul 2016 - Oct 2020

Head of Vision and Motion Group · Hangzhou

Designed a hierarchical reinforcement learning framework coordinating strategy and combat policies for complex multi-agent games. Invited to present the work at GTC.
Proposed a semi-supervised music-to-dance generation framework, using a multi-task music VAE and segment-level motion decoding to mitigate the scarcity of paired data.
Developed AI motion capture and expression retargeting, integrated into an art asset production platform supporting several large MMO titles.

Sep 2019 - Dec 2025

PhD in Artificial Intelligence, College of Computer Science · Hangzhou

Sep 2014 - Jul 2016

MSc in Automation and Intelligence · Beijing

Sep 2010 - Jul 2014

BSc in Automation · Beijing

Full list on the papers page. xiangwen-zju.com/papers

SkyReels-V1
Open-source human-centric video foundation model. Text-to-video and image-to-video weights plus the SkyReelsInfer inference framework.
SkyReels-A1
Expressive portrait animation with video diffusion transformers.
HY-WU
Functional neural memory framework from Tencent Hunyuan. Generates instance-conditioned LoRA adapters on the fly for text-guided image editing.

Business Application Pioneer Award2026
Tencent
For leading the algorithm work on AniMatrix, an anime video generation model, covering the data system, multi-stage post-training, and inference optimization.

Game Developers Conference, Machine Learning Summit2023
GPT-3 powered text to lifelike speech and animation for NPCs
Apsara Conference2024
AI workflows for producing drama online
Beijing International Game Innovation Conference (BIGC)2020
AI-assisted game content production
NVIDIA GTC China2019
Video-guided intelligent choreography

Two US patents on face processing and animation generation.