Scaling Rectified Flow Transformers for High-Resolution Image Synthesis

Diffusion models create data from noise by inverting the forward paths that turn data into noise, and rectified flow is a recent generative formulation that instead connects data and noise along a straight line, offering better theoretical properties and conceptual simplicity but not yet established as standard practice. In this work the authors improve existing noise sampling techniques for training rectified flow models by biasing them toward perceptually relevant scales, and they introduce a transformer-based text-to-image architecture that uses separate weights for the image and text modalities while enabling a bidirectional flow of information between image and text tokens.

Through a large-scale study, they demonstrate that this rectified flow approach outperforms established diffusion formulations for high-resolution text-to-image synthesis, while the two-stream transformer improves text comprehension, typography, and human preference ratings. The architecture follows predictable scaling trends, with lower validation loss correlating with improved text-to-image synthesis across various metrics and human evaluations, and the largest models outperform state-of-the-art systems; the authors state they will make their experimental data, code, and model weights publicly available.