Photographic Image Priors in the Era of Machine Learning - 2023
Details
Title : Photographic Image Priors in the Era of Machine Learning Author(s): MITCBMM Link(s) : https://www.youtube.com/watch?v=GK_0BqOX4QY
Rough Notes
I found interesting the idea where they rewrote the conditional mean estimate of the denoising (which is an integral) in terms of the gradient of the log density of the noisy images. The conditional mean values \(\hat{x}(y)\) are then taken from an already trained denoising network, then a sample \(y_t\) (which starts as noise) is generated by updating it with \(\hat{x}(y_{t-1})-y_{t-1}\) (which is equivalent to \(\nabla_y \log p(y)\) by the rewriting mentioned above).