Deep Equilibrium Approaches to Diffusion Models

Authors: Ashwini Pokle, Zhengyang Geng, J. Zico Kolter

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate our method s strong performance across several datasets, including CIFAR10, Celeb A, and LSUN Bedroom and Churches.1
Researcher Affiliation Collaboration Ashwini Pokle Carnegie Mellon University apokle@cs.cmu.edu Zhengyang Geng Carnegie Mellon University zgeng2@cs.cmu.edu Zico Kolter Carnegie Mellon University Bosch Center for AI zkolter@cs.cmu.edu
Pseudocode Yes Algorithm 1 A naive algorithm to invert DDIM ... Algorithm 2 Inverting DDIM with DEQ
Open Source Code Yes Code is available at https://github.com/ashwinipokle/deq-ddim
Open Datasets Yes We consider four datasets that have images of different resolutions for our experiments: CIFAR10 (32 32) [46], Celeb A (64 64) [52], LSUN Bedroom (256 256) and LSUN Outdoor Church (256 256) [76].
Dataset Splits No The paper mentions using well-known datasets (CIFAR10, Celeb A, LSUN) and refers to 'additional experimental details in the Appendix A', but the provided text does not explicitly state the training/validation/test splits, percentages, or sample counts used for these datasets.
Hardware Specification Yes All the experiments have been performed on NVIDIA RTX A6000 GPUs.
Software Dependencies No The paper mentions 'PyTorch-style pseudocode' and 'modern autograd packages' but does not specify exact version numbers for these or other software dependencies.
Experiment Setup Yes For all the experiments, we use Anderson acceleration as the default fixed point solver. While training DEQs for model inversion, we use the 1-step gradient Eq. (18) to compute the backward pass. The damping factor for 1-step gradient is set to 0.1. For the forward pass of DEQ, we run Anderson solver for a maximum of 15 steps for each image.