/Manifold-Constrained-Gradient-ipynb

Unofficial implementation for the paper 'Improving Diffusion Models for Inverse Problems using Manifold Constraints'[https://arxiv.org/abs/2206.00941]

Primary LanguageJupyter NotebookMIT LicenseMIT

MCG-ipynb

Unofficial implementation for inpainting and colorization experiments from the paper 'Improving Diffusion Models for Inverse Problems using Manifold Constraints'

Usage

To use these notebooks, first clone ilvr_adm repo and download pre-trained unconditional models FFHQ, then move these two notebooks into ilvr_adm folder and move ffhq_10m.pt into the path ./ilvr_adm/models.

In both notebooks, reconstruct_vanilla implement the imputation process discussed in the Score-SDE paper [1], while reconstruct_mcg refers to Manifold Constrained Gradient (MCG) method [2].

Results:

Grey:

a

Colorized using Score-SDE:

b

Colorized using MCG:

c

*I also did some audio spectral inpainting experiments in mel-scale trained with GradTTS backbones and Score-SDE based method: a

Reference

[1] Song, Yang, et al. "Score-Based Generative Modeling through Stochastic Differential Equations." International Conference on Learning Representations. 2021.

[2] Chung, Hyungjin, et al. "Improving Diffusion Models for Inverse Problems using Manifold Constraints." arXiv preprint arXiv:2206.00941 (2022).