Jiaxu Zhang | 张嘉旭

I am a research scientist at ByteDance Seed, focusing on AIGC and MLLM research.

I previously interned at Tencent from 2022 to 2024 and at StepFun from 2024 to 2025, where I worked closely with Dr. Gang Yu. I received my Ph.D. degree jointly from Wuhan University and Nanyang Technological University, under the supervision of Prof. Zhen Dong and Prof. Guosheng Lin.

My research interests cover computer vision and computer graphics, with a particular focus on multimodal AIGC, 2D/3D character animation, video/motion generation, retargeting, and recognition.

Feel free to reach out via WeChat if you are interested in research collaboration.

Email  /  CV  /  Google Scholar  /  Github  /  WeChat

We present FlowAct-R1, a novel framework that enables lifelike, responsive, and high-fidelity humanoid video generation for seamless real-time interaction.

UniMAGE unifies script drafting, extension, continuation, and keyframe image generation, thereby enabling coherent long-form storytelling with consistent characters and cinematic visual compositions. The generated scripts and keyframes can further serve as structured, high-level guidance for existing audio-video joint generation models.

We propose DreamDance, a novel paradigm that reformulates the character art animation task into two inpainting based steps: Camera-aware Scene Inpainting for stable scene reconstruction and Pose-aware Video Inpainting for dynamic character animation.

We propose MikuDance, a diffusion-based pipeline incorporating mixed motion dynamics to animate stylized character art.

A project for transforming Genshin PVs into a freehand style using Diffusion Model.

I've been exploring 2D image animation recently. This project is purely for fun. Feel free to reach out and discuss this with me.

M-R2ET is a modular neural motion retargeting system designed to transfer motion between characters with different structures but corresponding to homeomorphic graphs, meanwhile preserving motion semantics and perceiving shape geometries.

We present MotionS, a generative motion stylization pipeline for synthesizing diverse and stylized motion on cross-structure source using cross-modality style prompts.

TapMo is a text-based animation pipeline for generating motion in a wide variety of skeleton-free characters.

R2ET is a neural motion retargeting model that can preserve the source motion semantics and avoid interpenetration in the target motion.

We propose a novel Zoom Transformer to exploit both the low-level single-person motion information and the high-level multi-person interaction information in a uniform attention structure.

we propose a semi-supervised skeleton-based action recognition method.