Peng Chen

I am a third-year master student at Institute of Software, Chinese Academy of Sciences. I received my B.S. in Computer Science from University of Science and Technology, Beijing in 2023 and obtained Beijing Distinguished Graduate Award and Beijing Outstanding Graduation Thesis.

I serve as a reviewer for international conferences including ICLR, AAAI, ICME and ISMAR.

My research focuses on the following areas:

• MLLM (VLM/RL): video understanding, and GUI/Game/Embodied agent;
• AIGC (Diffusion/DiT): video/image generation, and unified model;
• 3D Vision (3DGS): digital humans;

I am currently looking for campus recruitment opportunities for 2026, focusing on multimodal large language models(MLLM).

Email  /  Github  /  Google Scholar

We propose CombatVLA, the first efficient visual-language action model designed for combat tasks in 3D action role-playing games. For efficient decision making, our CombatVLA is a 3B model that processes visual inputs and outputs a sequence of actions to control the game (including keyboard and mouse operations).

We propose a unified VLM named Visual-CoG, which leverages reinforcement learning (RL) with stage-aware rewards to provide immediate guidance throughout the image generation process, significantly improving performance on complex text-to-image tasks.

We use 2DGS to maintain the surface geometry and employ 3DGS for color correction in areas where the rendering quality of 2DGS is insufficient, reconstructing a realistically and geometrically accurate 3D head avatar.

RSATalker achieves realistic talking head generation for multi-turn conversation. It can perceive the social relationship between the speaker and listener, thereby expressing facial movements more accurately.

We propose GazeGaussian, a high-fidelity gaze redirection method that uses a two-stream 3DGS model to represent the face and eye regions separately.

We propose GraphAvatar, a compact method using Graph Neural Networks (GNN) to generate 3D Gaussians for head avatar animation, offering superior rendering performance and minimal storage requirements.

We propose a novel framework named the VARP agent, which directly takes game screenshots as input and generates keyboard and mouse operations to play the ARPG.

We propose DiffusionTalker, a diffusion-based method that utilizes contrastive personalizer to generate personalized 3D facial animation and personalizer-guided distillation for acceleration and compression.

Given a target facial video as reference, bring your own character into our solution integrated with Unity3D, it automatically generates facial animation for the virtual character.