How I Warped Your Noise (ICLR 2024) - Unofficial

An unofficial implementation of the ICLR 2024 paper "How I Warped Your Noise: a Temporally-Correlated Noise Prior for Diffusion Models", Chang et al. Implemented by Min-Seop Kwak.

Introduction

• The Jupyter Notebook file "How_I_Warped_Your_Noise_Unofficial" has been originally written for Google Colab. For direct usage at Colab, we provide the direct link to the Colab.

• For local usage, run it on a Conda environment that has einops, ninja, and Nvdiffrast installed.

Additional Details

Cell 9 contains the following code:

  warp_idxs = torch.stack((warp_i, warp_j), dim=-1)
  tgt_to_src_map = warp_idxs

Variable tgt_to_src_map designates the corresponding locations that each vertex of partitioned polygons of the target frame is mapped to at the source frame (warped locations).

• The current code provides simple warping configurations, identity mapping (no change), and rotation mapping (slight rotation), for testing purposes.
• Replace the variable with your desired correspondence mapping for your personal usage.

---

This implementation contains two separate implementations for triangle rasterization, at Cells 12 and 13.

• Using the code at Cell 12, which uses scatter torch.scatter_add_() is more efficient and therefore recommended.
• The code at Cell 13 uses for loop for sequential rasterization, and is therefore much slower, but computationally cheaper.

Citation

@inproceedings{
chang2024how,
title={How I Warped Your Noise: a Temporally-Correlated Noise Prior for Diffusion Models},
author={Pascal Chang and Jingwei Tang and Markus Gross and Vinicius C. Azevedo},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=pzElnMrgSD}
}