Accelerating Convergence of Score-Based Diffusion Models, Provably
Authors: Gen Li, Yu Huang, Timofey Efimov, Yuting Wei, Yuejie Chi, Yuxin Chen
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we illustrate the performance of the proposed accelerated samplers, focusing on emphasizing the relative comparisons with respect to the original DDIM/DDPM ones using the same pre-trained score functions. We specifically report results for deterministic samplers here, leaving the stochastic setting to Appendix A. Figure 1 illustrates the progress of the generated samples over different numbers of function evaluations (NFEs) (between 5 and 50) from the same random seed, using pretrained scores from the LSUN-Churches dataset. Figure 3. The FID of the DDIM-type samplers for different datasets with respect to the NFEs. |
| Researcher Affiliation | Academia | 1Department of Statistics, The Chinese University of Hong Kong, Hong Kong 2Department of Statistics and Data Science, Wharton School, University of Pennsylvania, Philadelphia, PA 19104, USA 3Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA. |
| Pseudocode | No | The paper describes algorithms and update rules within the text (e.g., 'the proposed discrete-time sampler adopts the following update rule: Y t = Φt(Yt), Yt 1 = Ψt(Yt, Y t )'), but it does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | The paper mentions utilizing pre-trained score functions from Huggingface (von Platen et al., 2022) and the DPM-Solver codebase (Lu et al., 2022a), but it does not provide an unambiguous statement or link for its own source code. |
| Open Datasets | Yes | We use pre-trained score functions from Huggingface (von Platen et al., 2022) for the Celeb A-HQ, LSUN-Bedroom, and LSUN-Churches datasets. Moreover, for the CIFAR-10 dataset, we utilize pre-trained score functions from Ho et al. (2020) and the DPM-Solver codebase (Lu et al., 2022a). For Image Net, we use the pre-trained score functions from Improved DDPM (Nichol and Dhariwal, 2021). |
| Dataset Splits | No | The paper does not provide specific dataset split information (e.g., percentages or sample counts) for training, validation, or test sets. It mentions using pre-trained models and evaluating over a range of NFEs. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper mentions using 'pre-trained score functions from Huggingface (von Platen et al., 2022)' and 'the DPM-Solver codebase (Lu et al., 2022a)', but it does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | No | The paper states: 'Note that we have not attempted to optimize the speed nor the performance using additional tricks, e.g., employing better score functions, but aim to corroborate our theoretical findings regarding the acceleration of the new samplers when the implementations are otherwise kept the same.' However, it does not provide specific hyperparameter values or detailed training configurations for the experimental setup beyond varying the Number of Function Evaluations (NFEs). |