Tweedie Moment Projected Diffusions for Inverse Problems

Authors: Benjamin Boys, Mark Girolami, Jakiw Pidstrigach, Sebastian Reich, Alan Mosca, Omer Deniz Akyildiz

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, Section 6 will present experiments on Gaussian mixtures, image inpainting and super-resolution, demonstrating quantitative and qualitative improvements provided by TMPD.
Researcher Affiliation Collaboration Benjamin Boys EMAIL Department of Engineering, University of Cambridge, Cambridge, United Kingdom; Alan Mosca EMAIL n Plan
Pseudocode Yes Algorithm 1 TMPD-D (Ancestral sampling, VP) input y, σy
Open Source Code Yes The code for all of the experiments and instructions to run them are available at github.com/bb515/tmpdjax and github.com/bb515/tmpdtorch.
Open Datasets Yes We consider inpainting and super-resolution problems on the FFHQ 256 256 (Karras et al., 2019) and CIFAR-10 32 32 (Krizhevsky et al., 2009) datasets.
Dataset Splits Yes We use a DDPM sampler, on FFHQ 256 256 using 1k validation images... We next compare performance to TMPD across VP and VE-SDE samplers and a range of noise levels on CIFAR-10 64 64 using 1k validation images... We chose the DPS scale hyperparameter by optimising LPIPS, MSE, PSNR and SSIM on a validation set of 128 images (see Fig. 9 for an example).
Hardware Specification Yes BB gratefully acknowledges the EPSRC for funding this research through the EPSRC Centre for Doctoral Training in Future Infrastructure and Built Environment: Resilience in a Changing World (EPSRC grant reference number EP/S02302X/1); and the support of n Plan, and in particular Damian Borowiec and Peter A. Zachares, for the invaluable facilitation of work that was completed whilst on internship with n Plan and access to A100 GPUs.
Software Dependencies No The paper mentions 'github.com/bb515/tmpdjax and github.com/bb515/tmpdtorch' which implies the use of JAX and PyTorch, but no specific version numbers for these or any other software dependencies are provided in the text.
Experiment Setup Yes We use 1000 timesteps for the time-discretization. For the Markov chain methods we use DDPM and for the SDE methods we use an Euler-Maruyama discretization... For super-resolution, we use a downsampling ratio of 4 (256 256 64 64) and bicubic interpolation; for box mask inpainting we mask out 128 128 region and for random mask inpainting we choose a random mask for each image masking between 30% and 70% of the pixels. Images are normalized to the range [0, 1] and it is on this scale that we add Gaussian measurement noise with standard deviation σy {0.01, 0.05, 0.1, 0.2}... we set the DDIM hyperparameter η = 1.0... We chose the DPS scale hyperparameter by optimising LPIPS, MSE, PSNR and SSIM on a validation set of 128 images (see Fig. 9 for an example). We found that static thresholding (clipping the denoised image estimate to a range [0, 1] at each sampling step) is critical for the stability and performance of both DPS-D and ΠGDM-D.