GENIE: Higher-Order Denoising Diffusion Solvers
Authors: Tim Dockhorn, Arash Vahdat, Karsten Kreis
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate GENIE on multiple image generation benchmarks and demonstrate that GENIE outperforms all previous solvers. Experimentally, we validate GENIE on multiple image modeling benchmarks and achieve state-of-the-art performance in solving the generative ODE of DDMs with few synthesis steps. |
| Researcher Affiliation | Collaboration | Tim Dockhorn1,2,3, Arash Vahdat1 Karsten Kreis1 1NVIDIA 2University of Waterloo 3Vector Institute tim.dockhorn@uwaterloo.ca, {avahdat,kkreis}@nvidia.com |
| Pseudocode | Yes | In App. C.2.4, we provide pseudo code for training and sampling with heads kψ. |
| Open Source Code | Yes | Project page and code: https://nv-tlabs.github.io/GENIE. |
| Open Datasets | Yes | Datasets: We run experiments on five datasets: CIFAR-10 [97] (resolution 32), LSUN Bedrooms [98] (128), LSUN Church-Outdoor [98] (128), (conditional) Image Net [99] (64), and AFHQv2 [100] (512). We also cite all used datasets: CIFAR-10 [97], LSUN Bedrooms [98], LSUN Church-Outdoor [98], Image Net [99], and AFHQv2 [100]. |
| Dataset Splits | No | The paper mentions running experiments on standard benchmark datasets like CIFAR-10 and Image Net, but it does not explicitly specify the train/validation/test dataset splits (e.g., percentages, sample counts, or explicit reference to predefined standard splits) within the provided text. While an ethics statement claims these details are in the Appendix, the main text itself is insufficient. |
| Hardware Specification | No | The paper states, 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] Please see App. F.8.', indicating that hardware specifications are located in the Appendix, but no specific hardware details (e.g., exact GPU/CPU models) are provided in the main text. |
| Software Dependencies | No | The paper does not provide specific ancillary software details, such as library or solver names with version numbers, that are needed to replicate the experiment. |
| Experiment Setup | Yes | Synthesis Strategy: We simulate the DDIM ODE from t=1 up to t=10 3 using evaluation times following a quadratic function (quadratic striding [58]). For variance-preserving DDMs, it can be beneficial to denoise the ODE solver output at the cutoff t=10 3... To this end, denoising the output of the ODE solver is left as a hyperparameter of our synthesis strategy. Analytical First Step (AFS)... AFS is treated as a hyperparameter of our Synthesis Strategy. Architectures... See App. C for training and architecture details. |