Unifying Bayesian Flow Networks and Diffusion Models through Stochastic Differential Equations

Authors: Kaiwen Xue, Yuhao Zhou, Shen Nie, Xu Min, Xiaolu Zhang, Jun Zhou, Chongxuan Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, using the same pre-trained model, our best solver significantly outperforms the original BFN sampler with a few (e.g., 10) number of function evaluations (NFE) under sample quality on both the CIFAR10 and text8 datasets, achieving a 5 20 times increase in speed for free (see Sec. 6 for details).
Researcher Affiliation Collaboration 1Gaoling School of AI, Renmin University of China, Beijing, China 2Department of Computer Science and Technology, Tsinghua University, Beijing, China 3Ant Group, Hangzhou, China.
Pseudocode Yes Algorithm 1 BFN-Solver++1 (on continuous data) ... Algorithm 2 BFN-Solver++2 (on continuous data) ... Algorithm 3 SDE-BFN-Solver++2 (on continuous data) ... Algorithm 4 SDE-BFN-Solver1 (on discrete data) ... Algorithm 5 SDE-BFN-Solver2 (on discrete data) ... Algorithm 6 BFN-Solver1 (on discrete data) ... Algorithm 7 BFN-Solver2 (on discrete data)
Open Source Code Yes Our code is available at https: //github.com/ML-GSAI/BFN-Solver.
Open Datasets Yes For continuous data, the model is trained on the CIFAR-10 (Krizhevsky & Hinton, 2009) dataset which contain 50K training images. For discrete data, the model is trained on the text8 (Mahoney, 2011) dataset which contains 90M consecutive characters, each character is a lower Latin letter a z or the whitespace token, giving a class number of 27.
Dataset Splits No The paper mentions training on CIFAR-10 and text8 datasets but does not explicitly provide details about training/validation/test splits (e.g., percentages, counts, or references to standard splits).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory) used for running the experiments.
Software Dependencies No The paper mentions using 'pre-trained models provided by the BFN (Graves et al., 2023)' but does not list specific software dependencies with version numbers.
Experiment Setup Yes We slightly tune the hyperparameter η for our methods on different NFEs to get the best results, as detailed in Appendix D.3.