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..
Improved Training Technique for Shortcut Models
Authors: Anh Nguyen, Viet Nguyen, Duc Vu, Trung Dao, Chi Tran, Toan Tran, Anh Tran
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Image Net 256 ˆ 256 demonstrate that our approach yields substantial FID improvements over baseline shortcut models across one-step, few-step, and multi-step generation, making shortcut models a viable and competitive class of generative models. |
| Researcher Affiliation | Industry | Qualcomm AI Research; Equal contribution :Work done while at Qualcomm. ;Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc. |
| Pseudocode | Yes | Pseudocode is provided in Appendix B. Algorithm 1 Multi-level Wavelet Function |
| Open Source Code | No | Yes, we will release our code upon acceptance. |
| Open Datasets | Yes | Extensive experiments on Image Net 256 ˆ 256 demonstrate that our approach yields substantial FID improvements over baseline shortcut models across one-step, few-step, and multi-step generation |
| Dataset Splits | Yes | For experiments on Image Net, images are preprocessed to 256x256 resolution, following the protocol of ADM [12]. We follow the ADM [12] evaluation setup, using the same reference batches from their official implementation,4 and compute FID [20] over 50K generated images. |
| Hardware Specification | No | The paper does not explicitly describe the hardware used for running its experiments. While the NeurIPS checklist states that this information is in the Appendix, the Appendix (specifically Table 6) does not contain specific hardware details like GPU or CPU models. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers. While it refers to architectures like Si T-XL in Table 6, it does not list programming languages, libraries, or frameworks with their respective versions. |
| Experiment Setup | Yes | Detailed configurations for all experiments are provided in Table 6. Table 6 includes 'Architecture', 'GFlops', 'Params (M)', 'Flow Trajectory', 'Input dim.', 'Num. layers', 'Hidden dim.', 'Num. heads', 'αt', 'σt', 'wt', 'Training objective', 'Training iteration', 'Dropout', 'Optimizer', 'Adam W β1', 'Adam W β2', 'Adam W ϵ', 'Learning Rate', 'Weight Decay', 'Batch Size', 'Label Dropout', 'CFG Scale wmax', 'Interval tinterval', 'Wavelet Levels L', 'OT Scale K', 'Ratio of Empirical to Self-consistency Targets', 'EMA Parameters Used For Self-consistency Targets?', 'EMA Target Rate θ target', 'EMA Parameters Used For Evaluation?', 'EMA Inference Rate θ infer'. |