LG AI Research: Research Trends in Score-based Generative Models at NeurIPS 2023

Recently, research on Score-based Generative Modeling, also known as Diffusion Models, has been actively progressing based on their outstanding performance. At NeurIPS 2023, various studies aimed at improving the model's structure and expanding its range of applications drew attention.

In terms of diffusion model architecture, setting the weighting during training is key. Google DeepMind improved performance through research combining the diffusion model's training loss with the ELBO (Evidence Lower Bound), which showed more consistent performance improvements than existing simple weighting methods.

Also notable is research aimed at increasing the model's efficiency and speed.

• U-Net Design Optimization: A multi-resolution U-Net structure was proposed that reduces the number of parameters while maintaining performance by utilizing basis functions such as Wavelets.
• SnapFusion: Through Data Distillation techniques, this dramatically increased sampling speed, enabling text-to-image generation in under 2 seconds even on mobile devices.

In the downstream tasks field, research using text-to-image diffusion models as Zero-Shot Classifiers beyond simple generation drew attention. According to research from Google DeepMind, by scoring the differences in generated results conditioned on prompts for specific classes, it was possible to achieve zero-shot classification performance stronger than existing CLIP models.