Yaru Liu

Hi there! I’m a third year PhD student in the Department of CST at University of Cambridge, advised by Prof. Rafal Mantiuk. My PhD research lies in the intersection field of Computer Graphics and Machine Learning. My research background includes content adaptive rendering and real-time 3D Reconstruction.

Previously, I obtained my B.Sc. degree in pure math at University of Toronto, and Msc. degree in computer engineering at McGill University, where I was gratefully supervised by Prof. Derek Nowrouzezahrai and Prof. Morgan McGuire.

Beyond research, I love business and fashion. I’ve created 3 startups with 1 exit, and am super interested in things with high technical barriers. I am driven by applicable questions, and sometimes use business to anchor my academic pursuits. Beyond that, fashion serves as my creative therapy. To me, these three don’t just coexist — they complement and elevate one another.

Feel free to reach out for collaboration🩵

News

Projects

Streaming of rendered content with adaptive frame rate and resolution
Yaru Liu, Joseph March, Rafal Mantiuk
Accepted to SIGGRAPH 2026

We exploit the spatio-temporal limits of the human visual system to adaptively adjust frame rate and resolution based on scene content and motion. A lightweight neural network predicts the optimal configuration to maximize perceptual quality while minimizing rendering load under bandwidth constraints. Our approach successfully optimizes perceived quality while reducing 50%+ computational power.
V-CAGE: Vision-Closed-Loop Agentic Generation Engine for Robotic Manipulation
Yaru Liu, Ao-bo Wang, Nanyang Ye, 2026

We present V-CAGE, an agentic framework for autonomous robotic data synthesis that leverages foundation models to bridge high-level semantic reasoning with low-level physical interaction. By centralizing semantic layout planning and visual self-verification, V-CAGE fully automates the end-to-end pipeline for highly scalable robotic datasets.
Seeing enough: non-reference perceptual resolution selection for power-efficient client-side rendering
Yaru Liu, Dayllon Vinícius Xavier Lemos, Ali Bozorgian, Chengxi Zeng, Alexander Hepburn, Arnau Raventos, 2026
[Paper]

We propose a non-reference method leveraging the spatio-temporal limits of human vision to predict the lowest resolution that remains perceptually indistinguishable from maximum quality. This enables highly efficient, perception-guided client-side rendering on power-constrained devices.
SPEM-F: Accounting for Eye Motion in Image and Video Quality Metrics
Pontus Ebelin, Yaru Liu, Niklas Sanden, Dounia Hammou, Daqi Lin, Tomas Akenine-Möller, Rafal Mantiuk, 2026

We present SPEM-F, a novel preprocessing filter that models smooth pursuit eye motion to convert any image or video metric into a perceptual video quality metric. Validated on a new 240-fps dataset, it physiologically simulates visual system latency, significantly improving prediction accuracy for temporal rendering artifacts and fast motion.
M3ashy: Multi-modal material synthesis via hyperdiffusion
Chenliang Zhou, Zheyuan Hu, Alejandro Sztrajman, Yancheng Cai, Yaru Liu, Cengiz Oztireli
Proceedings of the AAAI Conference on Artificial Intelligence, 2025
[Paper]

This framework enables neural material synthesis utilizing hyperdiffusion to learn the distribution over material weights. It provides flexible generation guided by multi-modal inputs such as material types, text descriptions, or reference images.
Real-Time Scene Reconstruction using Light Field Probes
Yaru Liu, Derek Nowrouzezahrai, Morgan McGuire
I3D 2024, Poster
[Paper]

Our approach leverages sparse real-world images to generate multi-scale implicit representations of scene geometries. By introducing a novel probe data structure, we accurately capture depths to decouple rendering performance from scene complexity.