Reproduce “Frequency-Time Diffusion with Neural Cellular Automata” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
John Kalkhof, Arlene Kühn, Yannik Frisch, Anirban Mukhopadhyay
Researchers at Darmstadt University of Technology developed Diff-NCA and FourierDiff-NCA, two Denoising Diffusion Models built upon Neural Cellular Automata (NCA) to generate high-resolution images. FourierDiff-NCA achieved an FID score of 43.86 with 1.1 million parameters on CelebA, outperforming larger UNet-based models while demonstrating scale-invariant synthesis and multi-task capabilities like inpainting and super-resolution.
Key insights
- Integrating Fourier-based diffusion into NCAs enables efficient single-step global information acquisition, overcoming the long-standing limitation of NCAs in propagating global knowledge over many iterations.
- Neural Cellular Automata can be successfully adapted as parameter-efficient backbones for Denoising Diffusion Models, outperforming larger UNet-based counterparts for specific tasks.
- The inherent scale-invariance and local processing capabilities of NCAs allow for robust out-of-distribution image synthesis, inpainting, and super-resolution without explicit retraining for varied image sizes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.