Scaling Rectified Flow Transformers for High-Resolution Image Synthesis

Rectified Flow is a generative modeling approach that connects data and noise in a straight line, and despite its theoretical advantages and simplicity, it has not yet fully established itself as a standard approach. Performance was improved by refining existing noise sampling techniques to weight perceptually relevant scales.

Additionally, a new Transformer-based architecture was introduced that uses separate weights for the two modalities of text and images and supports bidirectional information flow between them. This architecture shows improved results in text comprehension, typography, and human preference ratings.

The key achievements of this research are as follows:

• Demonstrated superior performance compared to existing Diffusion models in high-resolution text-to-image synthesis
• Confirmed predictable scaling trends and demonstrated a correlation between validation loss and synthesis performance
• Achieved performance surpassing state-of-the-art (SOTA) models