DyPE: Dynamic Position Extrapolation for Ultra High Resolution Diffusion
TL;DR
DyPE (Dynamic Position Extrapolation) enables pre-trained diffusion transformers to generate ultra-high-resolution images far beyond their training scale. It dynamically adjusts positional encodings during denoising to match evolving frequency content—achieving faithful 4K × 4K results without retraining or extra sampling cost.
Installation
Create a conda environment and install dependencies:
conda create -n dype python=3.10
conda activate dype
pip install -r requirements.txt
Usage
Generate ultra-high resolution images with DyPE using the run_dype.py script:
python run_dype.py --prompt "Your text prompt here"
Key Arguments:
| Argument | Default | Description |
|---|---|---|
--prompt |
Dark fantasy scene | Text prompt for image generation |
--height |
4096 | Image height in pixels |
--width |
4096 | Image width in pixels |
--steps |
28 | Number of inference steps |
--seed |
42 | Random seed for reproducibility |
--method |
yarn |
Position encoding method: yarn, ntk, or base |
--no_dype |
False | Disable DyPE (enabled by default) |
Examples:
# Generate 4K image with default settings (YARN + DyPE)
python run_dype.py --prompt "A serene mountain landscape at sunset"
# Use NTK method without DyPE
python run_dype.py --method ntk --no_dype --prompt "A futuristic city skyline"
# Baseline comparison (no position encoding modifications)
python run_dype.py --method base
Generated images will be saved to the outputs/ folder (created automatically).
Citation
If you find this useful for your research, please cite the following:
@misc{issachar2025dype,
title={DyPE: Dynamic Position Extrapolation for Ultra High Resolution Diffusion},
author={Noam Issachar and Guy Yariv and Sagie Benaim and Yossi Adi and Dani Lischinski and Raanan Fattal},
year={2025},
eprint={2510.20766},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2510.20766},
}