Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Beyond Scores: Proximal Diffusion Models
Authors: Zhenghan Fang, Mateo Diaz, Sam Buchanan, Jeremias Sulam
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments. We begin with an example in cases where the scores and proximals can be computed to arbitrary precision. We set p0 as a uniform distribution over a discrete set of points (dataset from [55]), and run score-based diffusion sampling, as well as (PDA) and (PDA-hybrid) (see Algorithm 1). ... MNIST. We now move to a real, simple case for digits [17], where we must train score-networks (through score matching) as well as our learned proximal networks for Prox DM. ... CIFAR10. We perform a similar experiment on the CIFAR10 dataset [37]. ... Celeb A-HQ (256 × 256). We further evaluate our approach for high-resolution image synthesis using the Celeb A-HQ dataset at 256 × 256 resolution [34]. |
| Researcher Affiliation | Academia | Zhenghan Fang Mathematical Institute for Data Science Johns Hopkins University EMAIL Mateo Díaz Mathematical Institute for Data Science Johns Hopkins University EMAIL Sam Buchanan Toyota Technological Institute at Chicago EMAIL Jeremias Sulam Mathematical Institute for Data Science Johns Hopkins University EMAIL |
| Pseudocode | Yes | Algorithm 1 Proximal Diffusion Model (Prox DM) Algorithm 2 Proximal Matching Training |
| Open Source Code | Yes | Code for reproducing the experiments in this work is available at https://github.com/ZhenghanFang/ProxDM. |
| Open Datasets | Yes | MNIST. We now move to a real, simple case for digits [17], where we must train score-networks (through score matching) as well as our learned proximal networks for Prox DM. ... CIFAR10. We perform a similar experiment on the CIFAR10 dataset [37]. ... Celeb A-HQ (256 × 256). We further evaluate our approach for high-resolution image synthesis using the Celeb A-HQ dataset at 256 × 256 resolution [34]. |
| Dataset Splits | Yes | For CIFAR-10, we adopt the same U-Net as in [27]. ... Score matching follows the standard setup in [27] with learning rate warm-up, gradient clipping, and uniform t sampling as introduced in [75]. ... For MNIST, we halve the number of filters in the U-Net, following [72]. |
| Hardware Specification | No | The paper states in its NeurIPS checklist that hardware details are in the appendix, but the appendix does not contain specific hardware details such as GPU/CPU models, memory, or processor types. It only generally refers to 'GPU time' in the checklist itself, not within the body of the paper. |
| Software Dependencies | No | The paper does not explicitly mention specific software dependencies with version numbers, such as Python or PyTorch versions, that would be needed to replicate the experiments. |
| Experiment Setup | Yes | The β(t) in forward process is set as linear, β(t) = βmin + (βmax − βmin) t, for t ∈ [0, 1], with βmin = 0.1, βmax = 20, as is standard [75]. Prox DM is pre-trained with an ℓ1 loss, followed by proximal matching loss with ζ = 1 in the first half and decayed to 0.5 in the second half. Both networks are trained for the same number of epochs (see more details in Appendix E.3). ... For CIFAR-10, we adopt the same U-Net architecture as in [27]. ... The Prox DM is trained for a total of 375k iterations with a batch size of 512. We use the ℓ1 loss during the first 75k iterations as a pretraining step, followed by proximal matching loss with ζ = 2 for 150k iterations, and then ζ = 1 for the last 150k iterations. The learning rate is set to 10^-4. For score models, we use a batch size of 128, and train for 1.5M iterations ... We set the learning rate to 2e-4 and use gradient clipping and learning rate warm-up, following [27, 75]. |