Yu Deng (邓誉)

Email: dengyu[at]microsoft.com       Google Scholar      Github

I am currently a senior researcher at Microsoft Research Asia (MSRA). My research interests include 3D vision&generation, neural rendering, and Embodied AI.

Prior to that, I was a researcher at Xiaobing.AI working on realistic virtual avatar generation. Before joining Xiaobing, I had received my Ph.D. degree from Tsinghua University under the supervision of Prof. Harry Shum in 2022, and worked closely with Jiaolong Yang and Xin Tong as a research intern at MSRA from 2017 to 2022. Before that, I had received my B.S. from Department of Physics in Tsinghua University in 2017.

We propose a native 3D generative model built on a unified Structured Latent representation and Rectified Flow Transformers, enabling versatile and high-quality 3D asset creation.

We propose a new advanced VLA architecture for robot manipulation, which leverages cognitive information extracted by powerful VLMs to guide action prediction of a specialized action module.

We present MoGe, a powerful model for recovering 3D geometry from monocular open-domain images.

We learn a lifelike 4D head synthesizer by creating pseudo multi-view videos from monocular ones as supervision.

We propose a one-shot 4D head synthesis approach for high-fidelity 4D head avatar reconstruction while trained on large-scale synthetic data.

We propose a novel approach that enables a 3D-aware GAN to generate images with both state-of-the-art photorealism and strict 3D consistency.

We propose GRAM-HD, a 3D-aware GAN that can generate photorealistic and 3D-consistent images at 1024x1024 resolution.

We propose a learning-based approach for high-fidelity and 3D-consistent novel view synthesis of monocular portrait images.

We propose a one-shot talking head synthesis approach with disentangled control over lip motion, eye gaze&blink, head pose, and emotional expression.

We propose AniFaceGAN, an animatable 3D-aware GAN for multiview consistent face animation generation.

We propose a generative model for synthesizing radiance fields of topology-varying objects with disentangled shape and appearance variations.

We propose Generative Radiance Manifolds (GRAM), a method that can generate 3D-consistent images with explicit camera control, trained on only unstructured 2D images.

We propose a novel Deformed Implicit Field (DIF) representation for modeling 3D shapes of a category and generating dense correspondences among shapes with structure variations.

We propose DiscoFaceGAN, an approach for face image generation of virtual people with disentangled, precisely-controllable latent representations for identity, expression, pose, and illumination.

We propose a learning-based approach for recovering the 3D geometry of human head from a single portrait image without any ground-truth 3D data.

We propose a novel deep 3D face reconstruction approach that leverages a robust hybrid loss function and performs multi-image face reconstruction by exploiting complementary information from different images for shape aggregation.

The website template was adapted from Jon Barron.