[NeurIPS 2023] Trends in Score-based Generative Modeling - LG AI Research BLOG

This post examines the latest research trends in Score-based Generative Modeling (Diffusion Model), which has been gaining significant attention recently. The major research can be broadly divided into model architecture improvements and downstream task expansion.

In terms of Diffusion Model Architecture, research aimed at improving training efficiency and performance is actively underway. Google DeepMind presented research addressing the weighting of Training Loss at each noise level, and the Monotonic Weighting design that follows up on NVIDIA's EDM research drew attention. In addition, research such as SnapFusion, which improves the U-Net structure to reduce parameters or uses Data Distillation to run quickly even on mobile devices, has also emerged.

On the Downstream Task side, the core focus is on research that utilizes generative models for various tasks beyond simple image generation. Google DeepMind unveiled research that uses a Text-to-Image model as a Zero-Shot Classifier, which demonstrated superior performance compared to existing CLIP models and robustness against Texture Bias.

Beyond this, the scope of Diffusion model applications is rapidly expanding, including visual goal generation for Reinforcement Learning (RL), Point-cloud Completion, and Uni-ControlNet, which integrates various conditions.