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..

Spectral Analysis of Diffusion Models with Application to Schedule Design

Authors: Roi Benita, Miki Elad, Joseph Keshet

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We turn to empirically validate the schedules obtained by solving the optimization problem, referred to as the spectral schedule or spectral recommendation. In addition, we examine spectral phenomena arising from the diffusion process and their relation to the schedule structure. ... Finally, we apply our method to publicly available datasets, including CIFAR-10 [18], AFHQv2 [6], MUSIC [22] and SC09 [33], examining the relation to heuristic choices in past work.
Researcher Affiliation Academia Roi Benita Department of Electrical and Computer Engineering Technion, Haifa, Israel EMAIL Michael Elad Department of Computer Science Technion, Haifa, Israel EMAIL Joseph Keshet Department of Electrical and Computer Engineering Technion, Haifa, Israel EMAIL
Pseudocode No The paper describes procedures and equations (e.g., equations 1, 2, 3, 4, 8, 9, 10, 12) for the diffusion process and its analysis, but does not present any formal pseudocode or algorithm block labeled as such.
Open Source Code No The code for solving the optimization problem is not yet publicly available; However, we provide a detailed description of the optimization procedure in Sections 4 and 6, which can be reproduced using standard tools such as the scipy package. Additionally, we include the resulting spectral schedule used for CIFAR-10 in Appendix J, and we refer to the relevant open-source model for reproducibility. The corresponding code will be released shortly.
Open Datasets Yes Finally, we apply our method to publicly available datasets, including CIFAR-10 [18], AFHQv2 [6], MUSIC [22] and SC09 [33], examining the relation to heuristic choices in past work.
Dataset Splits No The paper mentions using datasets like CIFAR-10, AFHQv2, MUSIC, and SC09, and describes generating 50,000 samples for evaluation and training models. However, it does not explicitly state the training, validation, or test splits used for the models trained or evaluated in the paper's experiments.
Hardware Specification No All computations were performed on a standard CPU.
Software Dependencies No For solving the resulting optimization problems, we have employed the Sequential Least SQuares Programming (SLSQP) method [17], a well-suited method for minimization problems with boundary conditions, and equality and inequality constraints. ... can be reproduced using standard tools such as the scipy package.
Experiment Setup No For the MUSIC dataset, we trained a model based on the architecture proposed in [16, 1], using a linear noise schedule with T = 1000 diffusion steps. Training was performed in an unconditional setting on raw waveforms with a batch size of 64. For the SC09 dataset, we adopted the architecture from Git Hub , and similarly trained it with a linear schedule of T = 1000 steps.