Export
Feng Liu
· 5 previous affiliations
- Portland State UniversitySince 2010
- Full Professor · Since 2010
- Adobe Systems (United States)Since 2022
- Principal Research Scientist · Since 2022
- Meta (United States)2020–2021
- Visiting Research Scientist · 2020–2021
- Google (United States)2017–2018
- Visiting Researcher · 2017–2018
- University of Wisconsin–Madison2004–2010
- PhD student · 2004–2010
19 other authors named Feng Liu: Feng Liu (The University of Melbourne · RIKEN · University of Technology Sydney · 334 papers), Feng Liu (Rutgers, The State University of New Jersey · Stevens Institute of Technology · 232 papers), Feng Liu (Shenzhen University · 35 papers), Feng Liu (Drexel University · 10 papers), Feng Liu (6 papers), and 14 more
Research Overview
Feng Liu focuses on computer vision, computational photography, and applied machine learning. His early work addressed video stabilization, image retargeting, and parallax-tolerant stitching to improve the perceptual quality of dynamic scenes. More recently, his research investigates efficient rendering techniques, perceptual similarity metrics, and controllable video generation using diffusion models.
Research Trajectory
Feng Liu, holding roles as a Professor at Portland State University and a Principal Research Scientist at Adobe Research, initially focused his research on techniques for adapting and stabilizing visual media. In his early work, he explored Automatic image retargeting with fisheye-view warping to adjust image compositions for different displays. To address the challenges of hand-held camera footage, he developed Content-preserving warps for 3D video stabilization, a method that simulates directed camera motion by recovering 3D camera paths and applying perceptually plausible transformations. He expanded on this foundation by introducing Subspace video stabilization, which enforces subspace constraints on feature trajectories to achieve smooth, visually stable views over long videos.
Moving into complex image compositing, he contributed significantly to the research direction of Resolving Parallax Induced Misalignment in Multi View Alignment. In Parallax-Tolerant Image Stitching, he and his collaborators combined homography and content-preserving warping to seamlessly blend images even when perfect global alignment is impossible due to depth variations. As virtual reality emerged as a new medium, he investigated user interfaces for 360° stereoscopic video. In Depth Conflict Reduction for Stereo VR Video Interfaces, his team proposed techniques to dynamically adjust UI depth and apply localized blurring. This work helped mitigate visual discomfort associated with the research area of Resolving the Vergence Accommodation Conflict in Stereoscopic Displays.
In subsequent years, his focus expanded to differentiable rendering, frame interpolation, and rendering efficiency. To improve optical flow-based synthesis, he co-developed Softmax Splatting for Video Frame Interpolation, addressing the challenge of mapping multiple pixels to the same target location during forward warping. He also led efforts to accelerate computationally expensive Monte Carlo rendering. His group proposed Fast Monte Carlo Rendering via Multi-Resolution Sampling, using a deep convolutional neural network to fuse low-resolution/high-sample and high-resolution/low-sample renderings. They further optimized this pipeline in Future Frame Synthesis for Fast Monte Carlo Rendering by predicting future frames to reduce the number of pixels requiring ray sampling, and in Auxiliary Features-Guided Super Resolution for Monte Carlo Rendering, which uses high-resolution auxiliary features like albedo and normals to guide the upscaling of low-resolution renders.
During this period, he also addressed the evaluation of visual quality. Recognizing that existing metrics are sensitive to imperceptible translations, he introduced a Shift-Tolerant Perceptual Similarity Metric that incorporates anti-aliasing and improved border handling to better align with human visual persistence. Similarly, his team developed A Perceptual Quality Metric for Video Frame Interpolation to capture the unique temporal artifacts generated by interpolation algorithms. Exploring single-image animation, he presented a framework for the 3D Ken Burns Effect from a Single Image that leverages depth prediction to synthesize virtual camera scans with parallax.
Most recently, his trajectory has shifted toward the efficiency and controllability of generative AI, particularly video diffusion and autoregressive models. Advancing work in the area of Structural Consistency for High Velocity Video Generation, his lab introduced Progressive Growing of Video Tokenizers for Temporally Compact Latent Spaces, achieving significant temporal compression to train text-to-video diffusion models more efficiently. Contributing to the broader area of The Performance Ceiling of Base Generative Models, he guided the development of MotionCanvas: Cinematic Shot Design with Controllable Image-to-Video Generation, which translates 2D motion signals into scene-aware camera and object movements without requiring 3D-annotated training data. He also participated in research on visual autoregressive models, helping to address compounding errors during inference in Rethinking Training Dynamics in Scale-wise Autoregressive Generation. Continuing the theme of architectural efficiency, he contributed to Diffusion Transformer-to-Mamba Distillation for High-Resolution Image Generation, which distills self-attention-based diffusion transformers into linear-complexity state-space models to reduce the computational burden of high-resolution synthesis.
Current Directions
The most recent papers in the evidence indicate a strong focus on the efficiency, controllability, and architectural optimization of generative models for visual media. Recent work investigates how to make high-resolution image and video generation more computationally tractable. This includes developing temporally compact latent spaces for video tokenizers and exploring distillation techniques that transition from quadratic-complexity diffusion transformers to linear-complexity state-space models like Mamba. Additionally, current research addresses the practical usability of these generative systems, designing frameworks for cinematic shot design that allow precise, scene-aware control over camera and object motion in image-to-video generation. He is also involved in refining the training dynamics of scale-wise autoregressive models to mitigate error propagation during inference, ensuring higher fidelity in coarse-to-fine visual generation.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake