Liying Lu

We synthesize realistic low-light noise images in a spectral sampling way with minimal data acquisition. Only one single noisy image and one single dark frame per ISO is needed. Our method neither relies on simplified parametric models nor on large sets of clean-noisy image pairs.

We generate realistic low-light noise images through a specially designed diffusion model. We analyze the characteristics of the generated noise, and show the denoising results trained with our synthetic data.

Combines structure-accurate 3D priors and texture-rich 2D priors in pretrained generative networks for blind face restoration under extreme conditions

We leverage in-the-wild 2D talking-head videos to train a 3D facial animation model.

By relaxing the immutable one-to-one constraint of the single image super-resolution task, we allow the estimated patches to dynamically seek the best supervision during training.

This paper addresses the task of video frame interpolation, leveraging Transformers to model long-range pixel correlations across video frames.

We present a generic iterative alignment module which employs a gradual refinement scheme for sub-alignments of video restoration.

An approach for recovering high-frequency details in low-quality images by utilizing external reference images. We accelerate the correspondence matching process and introduce an improved feature adaptation method.

A multi-scale architecture used to complete the sparse depth map, with a knowledge distillation method employed to achieve high performance and fast speed.