Denoising Diffusion Step-aware Models

Authors: Shuai Yang, Yukang Chen, Luozhou Wang, Shu Liu, Ying-Cong Chen

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluations demonstrate that DDSM achieves computational savings of 49% for CIFAR-10, 61% for Celeb A-HQ, 59% for LSUN-bedroom, 71% for AFHQ, and 76% for Image Net, all without compromising the generation quality.
Researcher Affiliation Collaboration Shuai Yang 1,3 Yukang Chen 2 Luozhou Wang 1,3 Shu Liu 4 Yingcong Chen 1,5 1HKUST(GZ) 2CUHK 3HKUST(GZ) Smart More Joint Lab 4Smart More 5HKUST
Pseudocode Yes Algorithm 1 DDSM Training; Algorithm 2 DDSM Evolutionary Searching
Open Source Code Yes Our code and models are available at https://github.com/EnVision-Research/DDSM.
Open Datasets Yes We conduct experiments on five image datasets on different domains, ranging from small scale to large scale. They are CIFAR10 (Krizhevsky et al., 2009), Celeb A-HQ (64x64, 128x128) (Liu et al., 2015), LSUN-bedroom (Yu et al., 2016), AFHQ (Choi et al., 2020), and Image Net (Deng et al., 2009).
Dataset Splits No No explicit statement providing percentages or sample counts for training, validation, and test splits was found, although training data is mentioned. For the CIFAR-10 dataset, we utilized the standard split of 50,000 images designated for training. For the Celeb A-HQ-64, AFHQ, and LSUN-bedroom datasets, we adhered to the typical training data splits.
Hardware Specification Yes The GPU latency is the time cost of generating one image with a single NVIDIA RTX 3090.
Software Dependencies No No specific version numbers for software dependencies (e.g., pymoo, pytorch) are provided, only the names of the tools and their reference papers.
Experiment Setup Yes Regarding the sizes of the sub-networks, we offer seven different options, corresponding to 2/8 of the original ADM’s width... For the search parameters, the process encompasses a total of 10 iterations, with each iteration involving a population of 50, and maintaining a mutation rate of 0.001... For CIFAR-10, we set the weight parameter to 0.1 to favor higher image quality, while for Celeb A, the FLOPs weight parameter is adjusted to 0.25.