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..

Adaptive Discretization for Consistency Models

Authors: Jiayu Bai, Zhanbo Feng, Zhijie Deng, TianQi Hou, Robert Qiu, Zenan Ling

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that ADCMs significantly improve the training efficiency of CMs, achieving superior generative performance with minimal training overhead on both CIFAR-10 and Image Net.
Researcher Affiliation Collaboration Jiayu Bai1, Zhanbo Feng2, Zhijie Deng2, Tianqi Hou3, Robert C. Qiu1, Zenan Ling1 1School of EIC, Huazhong University of Science and Technology 2School of Computer Science, Shanghai Jiao Tong University 3Huawei
Pseudocode Yes Algorithm 1 Adaptive Discretization for Consistency Models
Open Source Code Yes Code is available at https://github.com/ rainstonee/ADCM.
Open Datasets Yes unconditional CIFAR-10 [16] and class-conditional Image Net 64 64 [3], respectively.
Dataset Splits Yes unconditional CIFAR-10 [16] and class-conditional Image Net 64 64 [3], respectively. We evaluate the sample quality using FID [8]
Hardware Specification Yes Table 6: Hyperparameter Settings ... GPU types RTX3090 RTX3090 A100 GPU memory 24G 24G 40G Number of GPUs 1 8 4
Software Dependencies No No specific version numbers for software dependencies (e.g., Python, PyTorch, CUDA versions) are explicitly listed in the paper.
Experiment Setup Yes Appendix A.2 Hyperparameters: Batch Size and EMA. For unconditional CIFAR-10, we use a batch size of 128 with an EMA decay rate of 0.9999... Lagrange Multiplier λ. For unconditional CIFAR-10, we set λ = 0.01... Table 6: Hyperparameter Settings ... Dropout probability 30% 40% 50% Optimizer RAdam Adam Adam Learning rate schedule fixed square root square root Learning rate max 0.0001 0.001 0.0009 Pseudo-Huber c 0.03 0 0