Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models
Authors: Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, Zhihua Zhang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To demonstrate the effectiveness and universality of Diff-Instruct, we consider two scenarios: distilling pre-trained diffusion models and refining existing GAN models. The experiments on distilling pre-trained diffusion models show that Diff-Instruct results in state-of-the-art single-step diffusion-based models. The experiments on refining GAN models show that the Diff-Instruct can consistently improve the pre-trained generators of GAN models across various settings. 4 Experiments |
| Researcher Affiliation | Collaboration | Weijian Luo1 , Tianyang Hu2 , Shifeng Zhang2, Jiacheng Sun2, Zhenguo Li2, Zhihua Zhang1 1Peking University, 2Huawei Noah s Ark Lab Email: luoweijian@stu.pku.edu.cn. Corresponding to: Tianyang Hu (hutianyang1@huawei.com) |
| Pseudocode | Yes | Algorithm 1: Diff-Instruct Algorithm |
| Open Source Code | Yes | Our official code is released through https://github.com/pkulwj1994/diff_instruct. |
| Open Datasets | Yes | The experiments on distilling pre-trained diffusion models on the Image Net dataset of a resolution of 64 64... On the Image Net 64 64 dataset, our Diff-Instruct achieves state-of-the-art performance... The CIFAR-10 Dataset. online: http://www. cs. toronto. edu/kriz/cifar. html, 55, 2014. (referring to [36]) and Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248 255. Ieee, 2009. (referring to [11]) |
| Dataset Splits | No | The paper mentions using CIFAR-10 and Image Net datasets for training and evaluation but does not provide specific details on training, validation, and test splits (e.g., percentages or sample counts) within the main text or appendices for reproducibility. |
| Hardware Specification | Yes | The test was run on 2 Nvidia V100 GPUs with 128 batch size and Py Torch distributed data-parallel mechanism. |
| Software Dependencies | Yes | Test environment: Py Torch 1.12.1 and Torchvision 0.13.1, and Torch.distributed.parallel on 2 V100 GPUs. |
| Experiment Setup | Yes | For additional details about the generator s architecture, pre-trained models, and the hyper-parameters on our experimental setup, please refer to Appendix B.1. ... We put detailed hyper-parameters for distilling in Table 7. ... We put detailed hyper-parameters for each experiment in Table 9. |