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..
Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism
Authors: Kunyun Wang, Bohan Li, Kai Yu, Minyi Guo, Jieru Zhao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform comprehensive experiments in various diffusion-based models, demonstrating that Para Step is an effective and generalized method applicable to both vision and audio modalities. |
| Researcher Affiliation | Academia | Kunyun Wang School of Computer Science, Shanghai Jiao Tong University EMAIL Bohan Li School of Computer Science, Shanghai Jiao Tong University EMAIL Kai Yu School of Computer Science, Shanghai Jiao Tong University EMAIL Minyi Guo School of Computer Science, Shanghai Jiao Tong University EMAIL Jieru Zhao School of Computer Science, Shanghai Jiao Tong University EMAIL |
| Pseudocode | Yes | A Round-based implementation of Para Step Algorithm 1 Round-based implementation of Para Step |
| Open Source Code | Yes | Our code is available at https://github.com/sjtu-zhao-lab/Para Step. |
| Open Datasets | Yes | We use the MS-COCO 2017 [20] validation set for text-to-image model, VBench [12] for text-to-video and image-to-video models, and Audio Caps [14] for audio model, Audio LDM2-large. |
| Dataset Splits | Yes | We use the MS-COCO 2017 [20] validation set for text-to-image model, VBench [12] for text-to-video and image-to-video models, and Audio Caps [14] for audio model, Audio LDM2-large. |
| Hardware Specification | Yes | Hardware All experiments are conducted on a machine equipped with 8 NVIDIA 4090 GPUs (24GB each) [28], connected via PCIe Gen3. |
| Software Dependencies | No | The paper does not explicitly mention any software dependencies with specific version numbers. |
| Experiment Setup | Yes | The number of warm-up steps for Para Step is set to 1 for Audio LDM2-large, 5 for SD3, 5 for SVD, 13 for Cog Video X-2b, and 18 for Latte. The number of inference steps is set to 200 for Audio LDM2-large and 50 for all other models. |