Yanhao Ge

Vivo · Tongji University
3 previous affiliations
  • VivoSince 2022
    • Researcher · Since 2022
    • PhD student · Since 2024
    • MS student · 2012–2016
    • Undergrad student · 2008–2012
  • Tencent (China)2016–2022
    • Researcher · 2016–2022
27Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Yanhao Ge focuses on computer vision and generative artificial intelligence, with specific emphasis on high-fidelity facial synthesis, 3D face modeling, and image restoration. His early work addressed foundational vision tasks like pedestrian attribute recognition and pose estimation before transitioning to advanced neural rendering and identity preservation techniques. Currently, his research explores diffusion-based image editing, ultra-high-resolution synthesis, and the integration of physical dynamics with neural models.

Research Trajectory

Ge’s early research addressed core computer vision tasks through specialized neural network architectures and data augmentation strategies. In 2018, he and his collaborators introduced Raftnet: Extract Task-aware Features for Pedestrian Attribute Recognition, which utilized a novel block design to extract task-specific features while sharing representations across multiple tasks. Building on this foundation of improving feature extraction, his work expanded to address data scarcity in complex vision problems. With Adversarial Semantic Data Augmentation for Human Pose Estimation (2020), he participated in developing a generative approach to synthesize challenging poses and occlusions, pasting segmented body parts with varying semantic granularity to improve model robustness.

Around 2020, Ge’s focus shifted significantly toward generative facial synthesis and 3D modeling. He oversaw the development of SimSwap: An Efficient Framework For High Fidelity Face Swapping (2020), which proposed an Identity Injection Module to transfer source identities into target faces while preserving attributes like expression and gaze. This work contributed to the broader area of Balancing Identity Preservation and Structural Facial Editing. Following this, he explored unsupervised 3D face modeling from unconstrained images. In Learning To Aggregate and Personalize 3D Face From In-the-Wild Photo Collection (2021), he helped design a framework that disentangles identity-consistent features from scene-specific factors without relying on controlled environments.

This interest in robust 3D facial reconstruction continued with Learning to Restore 3D Face from In-the-Wild Degraded Images (2022), which tackled the challenge of recovering fine-grained 3D details from low-resolution inputs by extracting generative facial priors. That same year, he contributed to Physically-guided Disentangled Implicit Rendering for 3D Face Modeling (2022). This framework bridged the gap between traditional graphics renderers and neural rendering by explicitly adopting 3D shading and rasterizing modules to guide implicit representations, ensuring both 3D-aware comprehension and high-realism image formation.

By 2024, Ge’s research incorporated diffusion models and real-time rendering techniques. He collaborated on StyleMaster: Towards Flexible Stylized Image Generation with Diffusion Models (2024), which enhanced pretrained text-to-image models with a multi-source style embedder and dynamic attention adapter, advancing research directions such as Decoupling Style from Content in Diffusion Models. Concurrently, he presented RealTalk: Real-time and Realistic Audio-driven Face Generation with 3D Facial Prior-guided Identity Alignment Network (2024), a framework that leverages 3D facial priors to achieve high-fidelity, person-generic face generation at 30 frames per second. Expanding beyond facial synthesis, he also investigated the intersection of physics and machine learning in NeuMA: Neural Material Adaptor for Visual Grounding of Intrinsic Dynamics (2024). This framework combined rigid physics-based simulators with flexible neural adaptors and 3D Gaussian Splatting to infer physical properties from objects in motion, intersecting with the area of Augmenting Physical Models with Neural Residuals and Learning Universal Dynamics from Physical Interaction.

Most recently, Ge has focused on the efficiency and scale of diffusion-based image generation and restoration. He contributed to PostEdit: Posterior Sampling for Efficient Zero-Shot Image Editing (2025), which explored posterior sampling techniques to manipulate images without extensive retraining, contributing to the research direction of The Fidelity-Editability Trade-off in Image Editing. To address the limitations of upscaling in generative models, his work on UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (2025) introduced a massive dataset of images exceeding 3K resolution alongside a frequency-aware training strategy, aligning with work in areas like Non-Uniform Time-Step Scheduling for Diffusion Models. This trajectory culminated in LearnIR: Learnable Posterior Sampling for Real-World Image Restoration (2026), which applied learnable posterior sampling to recover clean images from complex, real-world degradations like haze and motion blur, bypassing the need for precise mathematical models of the damage.

Current Directions

The most recent papers in the evidence indicate a strong focus on advancing diffusion models for complex image synthesis, zero-shot editing, and real-world restoration tasks. Recent work investigates efficient zero-shot image editing through posterior sampling, aiming to manipulate images while preserving underlying structural integrity. Additionally, his research addresses the challenges of ultra-high-resolution image generation, utilizing large-scale datasets and frequency-aware training strategies to produce crisp, fine-grained details rather than relying on artifact-heavy upscaling. Another active area involves real-world image restoration, where he explores learnable posterior sampling to recover high-quality images from complex degradations without requiring explicit forward operators or precise mathematical models of the image damage.

Profile updated

See something wrong? Report a mistake