Stochastic Interpolants: A Unifying Framework for Flows and Diffusions
Authors: Michael Albergo, Nicholas M. Boffi, Eric Vanden-Eijnden
JMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, algorithmic aspects are discussed and the approach is illustrated on numerical examples. ... We provide numerical demonstrations in line with these recommendations in Section 7, and we conclude with some remarks in Section 8. ... 7. Numerical results. So far, we have been focused on the impact of α, β, and γ in (4.1) on the density ρ(t), which we illustrated analytically. In this section, we study examples where the drift coefficients must be learned over parametric function classes. In particular, we explore numerically the tradeoffs between generative models based on ODEs and SDEs, as well as the various design choices introduced in Sections 3, 4, and 6. In Section 7.1, we consider simple two-dimensional distributions that can be visualized easily. In Section 7.2, we consider high-dimensional Gaussian mixtures, where we can compare our learned models to analytical solutions. Finally in Section 7.3 we perform some experiments in image generation. |
| Researcher Affiliation | Academia | Michael S. Albergo EMAIL Center for Cosmology and Particle Physics New York University New York, NY 10012, USA; Nicholas M. Boffi EMAIL Courant Institute of Mathematical Sciences New York University New York, NY 10012, USA; Eric Vanden-Eijnden EMAIL Courant Institute of Mathematical Sciences New York University New York, NY 10012, USA |
| Pseudocode | Yes | Algorithm 1: Learning b with arbitrary ρ0 and ρ1. Algorithm 2: Learning ηz with arbitrary ρ0 and ρ1. Algorithm 3: Learning ηos z with Gaussian ρ0. Algorithm 4: Sampling general stochastic interpolants. Algorithm 5: Sampling spatially-linear one-sided interpolants with Gaussian ρ0. |
| Open Source Code | No | The paper does not contain an unambiguous statement that the authors are releasing code for the work described, nor does it provide a direct link to a code repository. |
| Open Datasets | Yes | on the 128 × 128 Oxford flowers dataset Nilsback and Zisserman (2006) |
| Dataset Splits | No | Table 3 lists '# Training point' for Image Net (1,281,167) and Flowers (315,123), indicating the size of the training set used. However, it does not provide explicit details about how the dataset was split into training, validation, or test sets, nor does it refer to a standard splitting methodology that can be used for reproduction. |
| Hardware Specification | No | Table 3 in Appendix C.1 mentions '# GPUs 2' and '# GPUs 4' for different experiments but does not specify the model or type of GPUs (e.g., NVIDIA A100, RTX 3090). No other specific hardware details like CPU models, memory, or cloud instance types are provided. |
| Software Dependencies | No | The paper mentions 'U-Net architecture originally proposed in Ho et al. (2020)', 'Adam optimizer', 'dopri5 solver', and 'Heun method' but does not provide specific version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | Appendix C.1 provides 'Experimental Specifications' in Table 3, detailing 'Batch Size', 'Training Steps', 'Hidden dim', 'Attention Resolution', 'Learning Rate (LR)', 'LR decay', 'U-Net dim mult', 'EMA decay rate', 'EMA start iteration', and specific ranges/values for 't0, tf' for different learning and sampling scenarios. Section 7.1 further specifies 'Feed forward neural networks of depth 4 and width 512', '7000 iterations on batches comprised of 25 draws from the base, 400 draws from the target, and 100 time slices', and 'learning rate was set to .002 and was dropped by a factor of 2 every 1500 iterations of training'. |