Jing Liu
Entrepreneur, Technnologist, Creator
The Goal for Computer Graphics is to Create the Best Possible Tool for the Imagination.
Latest News
🦊 Lynx: Towards High-Fidelity Personalized Video Generation
🔥 [ Get the Model ]
Two papers accepted at CVPR 2025 !
Roles
Tech Lead / Research Scientist
TikTok (2019 - Present)
Personalized Generative Models for Image & Video Creation.
Founding Engineer
Bellus3D (2015 - 2019) [Acquired by Meta]
High fidelity 3D face scanning on mobile phones.
Shipped Features
Publication
ID-Patch: Robust ID Association for Group Photo Personalization
CVPR 2025
Generating personalized group photos is hard: Faces often end up misplaced, distorted, or unrecognizable. Our ID-Patch method fixes this with a novel approach that ensures perfect positioning and high resemblance, setting a new standard for AI-driven photo generation.
Read Paper
COAP: Memory-Efficient Training with Correlation-Aware Gradient Projection
CVPR 2025
Training large AI models is memory-intensive, and existing low-rank methods often slow down or degrade performance. Our method, COAP, optimizes memory while preserving speed and accuracy: achieving up to 81% memory savings and 4x faster fine-tuning without compromise.
Read Paper
AgileAvatar: Stylized 3D Avatar Creation via Cascaded Domain Bridging
SIGGRAPH Asia 2022
Creating stylized 3D avatars from selfie photos is challenging due to domain gaps and complex parameter tuning. Our self-supervised method bridges this gap, generating high-quality avatars that preserve identity, outperforming previous work and rivaling manual creation.
Read Paper
AgileGAN: Stylizing Portraits by Inversion-Consistent Transfer Learning
SIGGRAPH 2021
Turning real photos into high-quality stylized portraits is tough, with existing methods often producing artifacts. Our AgileGAN framework overcomes this by learning style transfer more effectively, generating stunning 1024×1024 portraits with just ~100 examples and ~1 hour of training.
Read PaperBlog
Contact
Mind Pixels