ottogin/locality-diffusion-demo

Locality in Image Diffusion Models

Interactive demo for the paper "Locality in Image Diffusion Models Emerges from Data Statistics".

Select an analytical denoiser baseline and a dataset, then generate images comparing the analytical model against a trained UNet. The demo also shows the closest training images to each generated sample.

Models:

• —PCA Locality — locality-based denoiser using PCA projections
• —Optimal — optimal denoiser using FAISS-based softmax-weighted KNN
• —Wiener — classical Wiener filter denoiser
• —Nearest Dataset — nearest-neighbor baseline (L2 in pixel space)

Links:

• —Paper
• —GitHub