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..
Simple Distillation for One-Step Diffusion Models
Authors: Huaisheng Zhu, Teng Xiao, Shijie Zhou, Zhimeng Guo, Hangfan Zhang, Siyuan Xu, Vasant G Honavar
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present results of experiments which show that CED achieves performance comparable to that of representative baselines for distilling multi-step diffusion models while maintaining excellent memory efficiency. ... 5 Experiment ... We present the image generation results comparing various baseline methods on the CIFAR-10 and Image Net datasets in Tables 2 and 3, respectively. |
| Researcher Affiliation | Collaboration | Huaisheng Zhu1, Teng Xiao3,4, Shijie Zhou2, Zhimeng Guo1, Hangfan Zhang1, Siyuan Xu1, Vasant Honavar1 1Pennsylvania State University, 2 University at Buffalo 3 Allen Institute for AI (AI2) 4 University of Washington EMAIL |
| Pseudocode | Yes | Algorithm 1 Simple Distillation of One-step Diffusion Models Require: Teacher diffusion model ptrue 1: Initialize the one-step generator with the teacher s score network Gθinit( ) strue( ) 2: Sample x from ptrue(x | z) with z N(0, I) 3: Warm-up train with L = E(x, z) to obtain a ones-step generator Gθ 4: Get negative samples x from Gθ (z) and construct a triplet (x, x , z) for each z N(0, I) 5: for each training iteration do 6: Get a batch of samples (x, x , z) 7: Update one-step generator s parameters θ with LCED in Equation (14) with (x, x , z) 8: end for |
| Open Source Code | Yes | Answer: [Yes] Justification: We provide the experimental results in Section 5 and code in supplementary materials to reproduce our results. ... Answer: [Yes] Justification: We provide code and data in supplementary materials with sufficient instructions. |
| Open Datasets | Yes | Datasets and Models. To thoroughly evaluate CED, we use two representative benchmark datasets from prior works: CIFAR-10 (32 32) for unconditional generation [28] and Image Net (64 64) for conditional generation [8]. |
| Dataset Splits | No | The paper states: 'To thoroughly evaluate CED, we use two representative benchmark datasets from prior works: CIFAR-10 (32 32) for unconditional generation [28] and Image Net (64 64) for conditional generation [8].' and 'For FID, we generate 50k samples and compare them against the training set used by the EDM teacher model as the reference.' It does not explicitly mention the train/validation/test splits used for training the models. |
| Hardware Specification | Yes | Training is conducted on 2 NVIDIA A100 GPUs. |
| Software Dependencies | No | We implement CED on top of the EDM codebase [22]. The paper does not provide specific version numbers for the EDM codebase or any other software libraries used. |
| Experiment Setup | Yes | We use the Adam optimizer with an effective batch size of 512 for CIFAR-10 and 1024 for Image Net. ... We train at fixed square resolutions and use a learning rate 3e-5. Moreover, we perform 10000 steps warm-up training as demonstrated in Line 3 of Algorithm 1. ... we randomly sample negative samples for only 25% of the batch size compared to the positive samples to optimize the objective effectively. ... we use 35 sampling steps for the Cifar-10 and 79 steps for Image Net with Heun s 2nd-order method [22]. |