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..
Unlocking Dataset Distillation with Diffusion Models
Authors: Brian Moser, Federico Raue, Sebastian Palacio, Stanislav Frolov, Andreas Dengel
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments across numerous Image Net subsets demonstrate that LD3M significantly outperforms the state-of-the-art at 128 128 and 256 256 resolutions, achieving superior cross-architecture generalization, i.e., 4.8 percentage points (1 IPC) and 4.2 points (10 IPC), and faster distillation times. |
| Researcher Affiliation | Collaboration | 1 University of Kaiserslautern-Landau, Kaiserslautern 2 German Research Center for Artificial Intelligence (DFKI), Kaiserslautern 3 ABB AG, Mannheim |
| Pseudocode | Yes | Algorithm 1 Latent Dataset Distillation with Diffusion Models (LD3M) Input: randomly selected collection Xs, pre-trained encoder E, pre-trained decoder D, pre-trained denoiser ยตฮธ with frozen parameters ฮธ, noise levels ฯt. |
| Open Source Code | Yes | The code for LD3M is provided at https://github.com/Brian-Moser/prune_and_distill. |
| Open Datasets | Yes | We conduct extensive experiments on 10 diverse 10-class subsets of Image Net-1k [8] at 128 128 (IPC=1, IPC=10) and 256 256 (IPC=1) resolutions, as well as CIFAR-10. |
| Dataset Splits | Yes | Following standard protocol, we distill datasets using DC [42], DM [41], or MTT [4] and evaluate by training unseen architectures (Alex Net [22], VGG-11 [33], Res Net-18 [17], Vi T [11]) from scratch on the distilled set, reporting mean test accuracy ( std. dev.) over 5 runs. |
| Hardware Specification | Yes | All experiments were run on a workstation equipped with an NVIDIA RTX A6000 GPU (48 GB VRAM). |
| Software Dependencies | Yes | Our implementation uses Py Torch 1.10.1 with torchvision 0.11.2, and we build upon the GLa D library for dataset distillation with a generative prior. |
| Experiment Setup | Yes | C Hyper-Parameters for Distillation Algorithms LDM. For all our LDM experiments, we set the unconditional guidance scale to the default value of 3. For 128 128 images, we used max. time steps of 10, and for 256 256 images, we used 20. DC. We utilize a learning rate of 10 3 throughout our DC experiments to update the latent code representation and the conditioning information. DM. In every DM experiment, we adopt a learning rate of 10 2, applying it to updates of the latent code representation alongside the conditioning information. MTT. For MTT experiments, a uniform learning rate of 10 is applied to update the latent code representation and the conditioning information. We buffered 100 trajectories for expert training, each with 15 training epochs. We used Conv Net-5 and Instance Norm. During dataset distillation, we used three expert epochs, max. start epoch of 5 and 20 synthetic steps. |