Phase-Preserving Diffusion

Structure-Aligned Generation for Games, Videos, and Simulators

Paper | Project Page | Video | ComfyUI | Structured Noise

Phase-Preserving Diffusion (PPD) is a drop-in change to the diffusion process that preserves image phase while diffusing magnitude, enabling geometry-consistent re-rendering. It works with any diffusion model (SD 1.5, FLUX, Wan) without architectural modifications or additional parameters.

Structural information is encoded in the phase. By replacing standard Gaussian noise with frequency-selective structured (FSS) noise, PPD preserves low-frequency phase to maintain geometry while allowing high-frequency appearance variation, controlled by a single cutoff radius parameter r.

Installation

pip install -r requirements.txt
pip install git+https://github.com/zengxianyu/structured-noise

Download model weights and place them in models/ppd/. Example input images are also available there.

Inference

SD 1.5

PYTHONPATH=. python examples/image_synthesis/sd_text_to_image_ppd.py \
  --input_image dog.jpg \
  --radius 15 \
  --prompt "A high quality picture captured by a professional camera. Picture of a cute border collie" \
  --output output.png

FLUX.1-dev

PYTHONPATH=. python examples/flux/model_inference/FLUX.1-dev_ppd.py \
  --input_image test2.jpg \
  --prompt "$(cat test2.txt)" \
  --output output.png \
  --radius 30

Wan2.2-14b

PYTHONPATH=. python examples/wanvideo/model_inference/Wan2.2-I2V-A14B_ppd.py \
  --input_image output.png \
  --input_video test2.mp4 \
  --prompt "$(cat test2.txt)" \
  --radius 30 \
  --output output.mp4

Training

FLUX

PYTHONPATH=. bash examples/flux/model_training/lora/PPD-FLUX.1-dev.sh

Uses photo-concept-bucket by default.

Wan — see training scripts in examples/wanvideo/. Uses open-sora-pexels-subset by default.

Acknowledgements

This repo is largely based on DiffSynth-Studio. Refer to the original repo for additional training scripts and use cases.

Citation

@article{zeng2025neuralremaster,
  title   = {{NeuralRemaster}: Phase-Preserving Diffusion for Structure-Aligned Generation},
  author  = {Zeng, Yu and Ochoa, Charles and Zhou, Mingyuan and Patel, Vishal M and
             Guizilini, Vitor and McAllister, Rowan},
  journal = {arXiv preprint arXiv:2512.05106},
  year    = {2025}
}