Jingyun Liang

I am currently a Senior Research Scienist at Adobe Research, Seattle. Before this, I worked at Alibaba as an Algorithm Expert, after graduation from the Computer Vision Lab, ETH Zürich. During my PhD, I was under the supervision of Prof. Luc Van Gool and Prof. Radu Timofte. My research is focused on visual generation and low-level vision, such as image and video generation/ editing / restoration.

Email  /  Google Scholar  /  Github

I am looking for Research Scientist Interns on image and video generation/ editing/ restoration for Adobe Research, as well as academic collbrators. Please send your CV to me if you are interested 😇

• 2026-09: 💪🏻💪🏻💪🏻 I joined Adobe Reserch as a Senior Reserch Scientist.
• 2026-06: 👀👀👀 See our paper on exploring the 3D awareness of video diffusion models (MeshToken).
• 2026-04: 🎉🎉🎉 Our work on decomposed motion editing (RealisMotion) has been accepted by ICML2026.
• 2026-02: 🎉🎉🎉 Our work on autoregressive video generation (Lumos-1) has been accepted by ICLR2026.

Exploring the 3D awareness of video diffusion models by using mesh tokenization for render-free human motion control.

Decomposed human motion control for realistic video generation in world space coordinates.

Autoregressive (frame-by-frame) discrete (discrete diffusion within a frame) video generation.

Motion-guided consistent and controllable long video generation.

3D point cloud-guided camera motion control together with human motion controls.

Low-light image enhancement utilizing event camera signals for precise illumination estimation.

Unified, highly efficient, and scalable framework designed for image restoration tasks.

Unified framework for event-based frame interpolation and ad-hoc deblurring in unconstrained real-world environments.

Efficient Transformer architecture for image restoration via sharing key semantic features.

Generate the motion first, and then generate the video.

Event-based frame interpolation integrated with joint dynamic scene deblurring.

Leveraging denoising diffusion models for modular plug-and-play image restoration.

A comprehensive large-scale dataset created specifically for image restoration tasks.

A practical real-world video denoising model with impressive results on real-world videos.

An interpretable method based on alternating optimization. It takes low-framerate low-resolution blurry videos as inputs and can deal with video deblurring, frame interpolation and super-resolution problems in a joint way.

A reference-based image super-resolution Transformer with state-of-the-art performance. It consists of texture feature encoder module, reference-based deformable attention module and residual feature aggregation module.

An unfolded end-to-end event camera-based motion deblurring method; a High-Quality Blur (HQBlur) dataset for event camera-based deblurring.

A transformer-based model that jointly extracts, aligns, and fuses frame features at multiple scales; state-of-the-art performance in video SR/ deblurring/ denoising.

A practical real-world image denoising model with impressive results on real-world images.

A transformer-based model that jointly extracts, aligns, and fuses frame features at multiple scales; state-of-the-art performance in video SR/ deblurring/ denoising/ frame interpolation/ space-time video SR.

A transformer-based image restoration that allows content-based interactions and long-range dependency modelling; state-of-the-art performance on image SR, denoising and JPEG compression artifact reduction.

The first practical real-world degradation model for training real-world image SR models; impressive results on real-world images.

It learns a bijective mapping between HR and LR image pairs by modelling low and high-frequency components; a unified framework for image SR and image rescaling.

A kernel estimation network with mutual affine convolution for spatially variant kernels; state-of-the-art blind SR performance.

A normalizing flow-based kernel prior for kernel modeling; state-of-the-art performance in unsupervised blind SR.