Diffusion Schrödinger Bridge with Applications to Score-Based Generative Modeling

Authors: Valentin De Bortoli, James Thornton, Jeremy Heng, Arnaud Doucet

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments. We validate our methodology by generating image datasets such as MNIST and Celeb A. In particular, we show that using multiple steps of DSB always improve the generative model. We also show how DSB can be used to interpolate between two data distributions.Generative modeling. DSB is the first practical algorithm for approximating the solution to the SB problem in high dimension (d = 3072 for Celeb A). Whilst our implementation does not yet compete with state-of-the-art methods, we show promising results with fewer diffusion steps compared to initial SGMs (Song and Ermon, 2019) and demonstrate its performance on MNIST (Le Cun and Cortes, 2010) and Celeb A (Liu et al., 2015).
Researcher Affiliation Academia Valentin De Bortoli Department of Statistics, University of Oxford, UK James Thornton Department of Statistics, University of Oxford, UK Jeremy Heng ESSEC Business School, Singapore Arnaud Doucet Department of Statistics, University of Oxford, UK
Pseudocode Yes Algorithm 1 Diffusion Schrödinger Bridge
Open Source Code No The paper points to a project webpage ('https://vdeborto.github.io/publication/schrodinger_bridge/') but does not explicitly state that the source code for their methodology is released or provide a direct link to a code repository.
Open Datasets Yes We validate our methodology by generating image datasets such as MNIST and Celeb A. ... MNIST (Le Cun and Cortes, 2010) and Celeb A (Liu et al., 2015).
Dataset Splits No The paper specifies parameters like 'N = 20' or 'N = 50' diffusion steps and 'batch size of 128', but does not explicitly provide details about train/validation/test dataset splits (e.g., percentages, sample counts, or citations to predefined splits).
Hardware Specification No The paper states 'Computing resources were provided through the Google Cloud research credits programme,' but does not specify any particular hardware details such as GPU/CPU models or specific cloud instance types used for running experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies, such as programming languages, libraries, or frameworks used (e.g., Python, PyTorch, TensorFlow versions are not mentioned).
Experiment Setup Yes We train each DSB with a batch size of 128, N = 20 and γ = 1/40. We let N = 20 and γk = 0.01, i.e. T = 0.2. A reduced U-net architecture based on Nichol and Dhariwal (2021) is used to approximate Bn k and F n k.