# Impact of Noise on Diffusion Models B11902147 曹鈞皓 B11902020 簡彣諺 --- ## Related Work - Blue noise for diffusion models --- ## Contents - What is blue noise? - Why blue noise? - Our plan <!-- - Generating blue noise is time consuming. - Analyze performance of different noises - Our plan --> --- ## What is blue noise? ---- |  |  | | ---- | ---- | | White Noise | Blue Noise | - Blue noise has no energy in the low-frequency region and consists only of high-frequency components. --- ## Why blue noise? ----  - Diffusion models generate data in a **coarse-to-fine** manner and have a hidden relationship with frequency components. ----  - Use a time-varying blended noise which shifts from Gaussian white noise to blue noise as the time step $t$ decreases. --- <!-- ## Generating blue noise is time consuming. ---- $LL^T=\Sigma$  - To create a noise mask with a specified covariance matrix $\Sigma$, apply Cholesky decomposition to generate blue noise on the fly. ---- - Padding/Tiling --- --> ## Analyze performance of different noises ----  - FID : compares the distribution of features between the generated images and real images. - Precision: fall within the distribution of real images - Recall: diversity of the generated images --- ## Our Plan ---- - Try violet noise(density $\propto f^2$) - Try combining red and blue noise, as well as exploring other combinations.