DyPE

Overview

DyPE (Dynamic Position Extrapolation) is a training-free method designed to extend the resolution capabilities of pre-trained diffusion transformer models for image generation. By dynamically adjusting positional encodings during the denoising process, DyPE allows models to generate images at resolutions far beyond their original training scale—such as 4K × 4K or higher—without requiring additional training or increased sampling costs. This approach is particularly effective for models like FLUX, enabling them to maintain high fidelity and detail at ultra-high resolutions.