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..
One-Step Offline Distillation of Diffusion-based Models via Koopman Modeling
Authors: Nimrod Berman, Ilan Naiman, Moshe Eliasof, Hedi Zisling, Omri Azencot
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | KDM achieves highly competitive performance across standard offline distillation benchmarks... We conduct a comprehensive unconditional and conditional evaluation showing that our method outperforms prior offline distillation approaches, achieving FID improvement, while enabling efficient, single-step generation. (From Abstract and Introduction) and Section 6: Experiments. |
| Researcher Affiliation | Academia | 1Ben-Gurion University of the Negev 2University of Cambridge |
| Pseudocode | Yes | Additionally, A pseudo-code of our KDM is provided in App. E.1. |
| Open Source Code | Yes | Our code is in https://github.com/azencot-group/KDM. |
| Open Datasets | Yes | We follow the protocol of GET [24] for generating noisy-clean pairs and for evaluation, using the CIFAR-10 [43], FFHQ 64 64 [37], and AFHQv2 64 64 [15] datasets. |
| Dataset Splits | No | We follow the protocol of GET [24] for generating noisy-clean pairs and for evaluation, using the CIFAR-10 [43], FFHQ 64 64 [37], and AFHQv2 64 64 [15] datasets. Our training setup matches that of PD [68] and GET [24]. For evaluation, we report image quality using Fréchet Inception Distance (FID) and Inception Score (IS), computed over 50k samples. |
| Hardware Specification | Yes | RTX 4090 GPU used with batch size 128. (Table 4) and All runs are on a single NVIDIA A6000 GPU or RTX4090. (Appendix E.3) |
| Software Dependencies | No | We train our models for 800k iterations with the Adam optimizer [39] at a fixed learning rate of 3e-4. |
| Experiment Setup | Yes | We train our models for 800k iterations with the Adam optimizer [39] at a fixed learning rate of 3e-4. We do not use warm-up, weight decay, or any learning-rate schedule... Table 7: Model Hyperparameters Across Datasets (Batch Size, Learning Rate, Iterations, Unet Out Channels, Unet Model Channels, Noisy Latent Injection, Adversarial Weight, Linear Module, Latent Dim Size are provided for Checkerboard, CIFAR-10, FFHQ, AFHQv2). |