Conditional Generation

Conditional generation focuses on creating new data instances that satisfy specific constraints or conditions, aiming to control the output of generative models. Current research emphasizes training-free methods, leveraging pre-trained models like diffusion models and transformers, and exploring novel architectures such as Schrödinger bridge-based approaches and idempotent networks to improve sample quality and efficiency. This field is significant for diverse applications, including drug discovery (molecule generation), data augmentation (for low-resource scenarios), and realistic data synthesis (e.g., time series, images, and audio), ultimately advancing scientific understanding and practical capabilities across various domains.