Deep Generative Model

less than 1 minute read

Controllable Image Generation with Lightweight Diffusion Models Building on our past success in creating and manipulating novel hairstyles with StyleGAN2, our research has now advanced to the next generation of generative models. We are currently focused on developing lightweight, diffusion-based models for controllable image generation. Our primary goal is to create highly efficient frameworks that enable intuitive, real-time image manipulation through simple user inputs, such as sketches or text prompts, while maintaining high-fidelity results. This research aims to make advanced content creation more accessible and interactive.
Real-time 3D Scene Reconstruction with Gaussian Splatting In the realm of 3D scene representation, our focus has evolved towards the cutting-edge technique of 3D Gaussian Splatting. We are actively researching methods to achieve real-time, high-fidelity generation of large-scale indoor maps. Our work concentrates on optimizing the capture-to-render pipeline, enabling instant reconstruction and interactive exploration of complex environments. This technology is being developed for seamless integration into robotics, augmented reality (AR), and digital twin applications.