Accelerating Parallel Sampling of Diffusion Models

Authors: Zhiwei Tang, Jiasheng Tang, Hao Luo, Fan Wang, Tsung-Hui Chang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments demonstrate that Para TAA can decrease the inference steps required by common sequential sampling algorithms such as DDIM and DDPM by a factor of 4 14 times. Notably, when applying Para TAA with 100 steps DDIM for Stable Diffusion, a widely-used text-to-image diffusion model, it can produce the same images as the sequential sampling in only 7 inference steps. The code is available at https://github.com/ TZW1998/Para TAA-Diffusion.
Researcher Affiliation Collaboration 1School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, China 2DAMO Academy, Alibaba Group 3Hupan Lab, Zhejiang Province 4Shenzhen Research Institute of Big Data, Shenzhen, China.
Pseudocode Yes Algorithm 1 Para TAA: Parallel Sampling of Diffusion Models with Triangular Anderson Acceleration
Open Source Code Yes The code is available at https://github.com/ TZW1998/Para TAA-Diffusion.
Open Datasets Yes In this section, we present the effectiveness of our approach in accelerating the sampling process for two prevalent diffusion models: Di T (Peebles & Xie, 2023), a class-conditioned diffusion model trained on the Imagenet dataset at a resolution of 256x256, and text-conditioned Stable Diffusion v1.5 (SD) (Rombach et al., 2022) with a resolution of 512x512.
Dataset Splits No The paper mentions using well-known models (DiT, Stable Diffusion) trained on datasets like Imagenet, but does not explicitly state the training, validation, or testing splits used for its own experiments or how data was partitioned for reproduction purposes.
Hardware Specification Yes We run these experiments using 8 A800 GPUs, each with 80GB of memory.
Software Dependencies No The paper mentions specific models like Stable Diffusion v1.5 but does not provide explicit version numbers for software libraries or dependencies such as PyTorch, TensorFlow, or specific CUDA versions.
Experiment Setup Yes For all algorithms, we use the same stopping threshold εt = τ 2g2(t)d with τ = 10 3, and initialize all variables with standard Gaussian Distribution. In all scenarios, we employ classifier-free guidance (Ho & Salimans, 2022) with a guidance scale of 5. Para TAA has two hyperparameters: the history size m and the order k, both of which are chosen via grid search.