Large-scale Reinforcement Learning for Diffusion Models
This addresses the issue of suboptimal and biased image generation in diffusion models, which is important for users relying on these models for ethical and high-quality outputs, though it is incremental as it builds on existing methods.
The paper tackles the problem of implicit biases in text-to-image diffusion models by using large-scale reinforcement learning to align them with human preferences, resulting in generated samples preferred by humans 80.3% of the time over the base model.
Text-to-image diffusion models are a class of deep generative models that have demonstrated an impressive capacity for high-quality image generation. However, these models are susceptible to implicit biases that arise from web-scale text-image training pairs and may inaccurately model aspects of images we care about. This can result in suboptimal samples, model bias, and images that do not align with human ethics and preferences. In this paper, we present an effective scalable algorithm to improve diffusion models using Reinforcement Learning (RL) across a diverse set of reward functions, such as human preference, compositionality, and fairness over millions of images. We illustrate how our approach substantially outperforms existing methods for aligning diffusion models with human preferences. We further illustrate how this substantially improves pretrained Stable Diffusion (SD) models, generating samples that are preferred by humans 80.3% of the time over those from the base SD model while simultaneously improving both the composition and diversity of generated samples.