On Inference Stability for Diffusion Models

Authors: Viet Nguyen, Giang Vu, Tung Nguyen Thanh, Khoat Than, Toan Tran

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on several benchmark datasets including CIFAR10, Celeb A, and Celeb A-HQ consistently show a remarkable improvement of our proposed method regarding the image generalization quality measured by FID and Inception Score compared to several DPM baselines.
Researcher Affiliation Collaboration Viet Nguyen1* , Giang Vu2,3*, Tung Nguyen Thanh2,3, Khoat Than2 , Toan Tran1 1Vin AI Research, Vietnam 2Hanoi University of Science and Technology, Vietnam 3 Viettel Group, Vietnam
Pseudocode Yes Algorithm 1: Conventional training Algorithm 2: Sampling Algorithm 3: Sequence-aware training
Open Source Code Yes Our code and pre-trained checkpoints are available at https://github.com/Vin AIResearch/SA-DPM.
Open Datasets Yes CIFAR10 32 32 (Krizhevsky 2012), Celeb A 64 64 (Liu et al. 2015) and one higher-resolution dataset Celeb A-HQ 256 256 (Karras et al. 2018).
Dataset Splits No The paper does not provide specific details on training/validation/test dataset splits, percentages, or sample counts.
Hardware Specification No The paper does not explicitly describe the specific hardware used for running its experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers needed to replicate the experiment.
Experiment Setup Yes The sampling timesteps for all the datasets are set to {10, 50, 100, 200, 1000}. ... we used the SA-2-DPM with the weight λ of Lsa set to 1... λ {0.5, 1, 2} for SA-2-DPM, λ {0.3, 0.6, 1.5} for SA-3-DPM and λ {0.2, 0.4} for SA-4-DPM.