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..
Kuramoto Orientation Diffusion Models
Authors: Yue Song, Andy Keller, Sevan Brodjian, Takeru Miyato, Yisong Yue, Pietro Perona, Max Welling
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on orientation-rich datasets, such as fingerprints, textures, and terrains, demonstrate that our Kuramoto orientation diffusion model consistently produces higher-fidelity images compared to standard diffusion baselines, often requiring fewer diffusion steps. Furthermore, our method remains competitive even on general CIFAR-10 benchmarks. Overall, this work bridges neural oscillation theory and modern score-based generative models, underscoring the potential of biologically inspired synchronization dynamics as structured priors. |
| Researcher Affiliation | Collaboration | Yue Song1, T. Anderson Keller2, Sevan Brodjian1, Takeru Miyato3,4, Yisong Yue1, Pietro Perona1, and Max Welling4,5 1 Caltech 2 Harvard University 3 University of Tübingen 4 University of Amsterdam 5 Cusp AI |
| Pseudocode | Yes | Algorithm 1 Training algorithm for Kuramoto orientation diffusion models. Algorithm 2 Inference algorithm for Kuramoto orientation diffusion models. |
| Open Source Code | Yes | Code is available at:https://github.com/King James Song/Orientation Diffusion. |
| Open Datasets | Yes | For standard benchmarking, we first test on CIFAR10 [35]. To specifically assess performance on orientation-dense data, we then apply our model to the SOCOFing fingerprint dataset [56], the Brodatz texture dataset [1, 8], and the ground terrain dataset [68]. ... In the Supplementary, we extend evaluation beyond images to (i) Earth and climate science datasets on the 2D Sphere [46, 45, 5, 18], and (ii) Navier-Stokes fluid velocity fields [6]. |
| Dataset Splits | No | The input image resolutions are 3 32 32 for CIFAR10 and 1 96 96 for SOCOFing. The Brodatz texture dataset originally consists of high-resolution samples depicting various textures (e.g., grass, water, sand, wool) with sizes of 1 512 512 or 1 1024 1024. Due to the limited number of available images, we increase the dataset size by dividing these high-resolution textures into smaller patches of 1 32 32. |
| Hardware Specification | Yes | A single NVIDIA A100 GPU is used for all the training and inference processes. |
| Software Dependencies | No | We use Adam W optimizer with a learning rate of 1e 4 and apply exponential moving average (EMA) updates with decay rate 0.995. |
| Experiment Setup | Yes | We use Adam W optimizer with a learning rate of 1e 4 and apply exponential moving average (EMA) updates with decay rate 0.995. Table 5 summarizes the linear schedules for the noise variance 2Dt, internal coupling strength K, and reference phase coupling Kref. We find that using K=3 and M=5 samples per step provides a good balance between training stability and computational efficiency. |