Restart Sampling for Improving Generative Processes
Authors: Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, Tommi Jaakkola
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentally, Restart consistently surpasses previous ODE and SDE solvers in both quality and speed over a range of NFEs, datasets, and pre-trained models. Our experiments are consistent with the theoretical bounds and highlight Restart s superior performance compared to previous samplers on standard benchmarks in terms of both quality and speed. |
| Researcher Affiliation | Collaboration | Yilun Xu MIT ylxu@mit.edu Mingyang Deng MIT dengm@mit.edu Xiang Cheng MIT chengx@mit.edu Yonglong Tian Google Research yonglong@google.com Ziming Liu MIT zmliu@mit.edu Tommi Jaakkola MIT tommi@csail.mit.edu |
| Pseudocode | Yes | Detailed pseudocode for the Restart sampling process can be found in Algorithm 2, Appendix B.2. Algorithm 3 One Step_Heun(sθ, xti, ti, ti+1) |
| Open Source Code | Yes | Code is available at https://github.com/Newbeeer/diffusion_restart_sampling |
| Open Datasets | Yes | datasets, and pre-trained models. Specifically, Restart accelerates the previous best-performing SDEs by 10 fewer steps for the same FID score on CIFAR-10 using VP [23] (2 fewer steps on Image Net 64 64 with EDM [13]) We further apply Restart to the text-to-image Stable Diffusion v1.5 pre-trained on LAION-5B [21] at a resolution of 512 512. We employ the commonly used classifier-free guidance [8, 20] for sampling, wherein each sampling step entails two function evaluations the conditional and unconditional predictions. Following [18, 20], we use the COCO [15] validation set for evaluation. |
| Dataset Splits | No | No explicit train/validation/test split percentages, sample counts for each split, or references to predefined splits with specific citations detailing the split methodology were found for all datasets used. While 'COCO [15] validation set for evaluation' is mentioned, specific split details for all datasets are not provided for reproducibility. |
| Hardware Specification | Yes | All the experiments are conducted on eight NVIDIA A100 GPUs. |
| Software Dependencies | No | The paper mentions using 'diffuser' code repository and 'DDIM and DDPM samplers implemented in the repo' but does not provide specific version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | For the proposed Restart sampler, the hyperparameters include the number of steps in the main/Restart backward processes, the number of Restart iteration K, as well as the time interval [tmin, tmax]. We provide the detailed Restart configurations in Appendix C.2. We employ a common MLP architecture with a latent size of 64 to learn the score function. The training method is adapted from [13], which includes the preconditioning technique and denoising score-matching loss [25]. |