We propose SnapGen, the first text-to-image model (379M) that can synthesize high-resolution images (1024x1024) on mobile devices in 1.4s, and achieve 0.66 on GenEval metric.

We introduce AsCAN, a hybrid neural network with asymmetric convolutional and transformer blocks, offering superior performance and efficiency across image recognition and generation tasks.

We clean the widely-adopted RefCOCO,+,g benchmarks and introduce Ref-L4, a New REC benchmark in the LMM Era.

We propose StableKD, a simple and efficient Knowledge Distillation framework that attains higher accuracy using fewer training epochs and less data.

We propose a simple yet fast and effective partial convolution (PConv), as well as a latency-efficient family of architectures called FasterNet.

TVConv works more computation-efficient than regular convolution when dealing with layout-specific tasks, e.g., face recognition.

We consider the ground truth uncertainty for joint demosaicking and denoising in the wild, which provides better restoration result and interpretability.

We propose TASFAR, a novel target-agnostic source-free domain adaptation method for regression tasks.

We propose CP-NeRF to enable training a one-for-all NeRF across diverse scenes.

We design a floor identification system for crowdsourced RF signals in a building using only one labeled data sample from the bottom floor.

We develop a practical geofencing system, solely based on ambient radio frequency (RF) signals, to enable applications like elderly care, dementia antiwandering, pandemic control, etc.