Techniques used to improve generative models' alignment with human preferences include reinforcement learning, preference optimization strategies like Diffusion-DPO and SFT, and models such as Stable Diffusion XL (SDXL). Additionally, frameworks like Kahneman-Tversky Optimization (KTO) have been adapted for text-to-image diffusion models.
Diffusion models are primarily designed for handling high-dimensional data like images and audio, and their applications span various domains such as art creation and medical imaging. These generative models aim to capture the underlying distribution of a given dataset and learn to generate new samples that resemble the training data through an iterative refinement process5.
Diffusion models, a type of generative model, have shown remarkable performance in generating synthetic data, including images and audio. In art creation, they can generate novel and creative pieces based on given prompts or styles. In medical imaging, diffusion models can synthesize realistic 3D medical data, helping to augment datasets, preserve privacy, and aid diagnosis and treatment planning.