Diffusion models have established themselves as a potent tool for generating high-dimensional data, such as images and videos. They perform this feat by reversing the progression of data transforming into noise. This article on arXiv presents progress on rectified flow models, which linearly connect data and noise, to improve noise sampling during training and ultimately enhance text-to-image synthesis’ quality. Noteworthy points include:

• Innovations in noise sampling techniques that prioritize perceptual scales.
• A new transformer architecture that allows a bidirectional flow of information between image and text tokens.
• The ability to follow predictable scaling trends, reducing validation loss, and increasing quality based on metrics and human assessments.
• The proposed models surpass current state-of-the-art in text-to-image synthesis.

This study signals a significant step toward more effective and precision-oriented image generation, leveraging the strengths of transformers and diffusion models.