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..
SCoT: Unifying Consistency Models and Rectified Flows via Straight-Consistent Trajectories
Authors: zhangkai wu, Xuhui Fan, Hongyu Wu, Longbing Cao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results demonstrate the effectiveness and efficiency of SCo T. ... We evaluate SCo T on both CIFAR-10 and Image Net... Table 2: Comparison of N-step (NFE) generation performance across diffusion models on CIFAR-10... Table 3: Comparison of N-step generation performance by different DMs on Image Net... |
| Researcher Affiliation | Academia | Zhangkai Wu School of Computing Macquarie University EMAIL Xuhui Fan School of Computing Macquarie University EMAIL Hongyu Wu School of Computing Macquarie University EMAIL Longbing Cao School of Computing Macquarie University EMAIL |
| Pseudocode | Yes | Regarding sampling, Algorithm 1 outlines the SCo T sample generation process, which may be implemented as a multi-step or single step procedure. ... Algorithm 1 Sampling Procedure of SCo T |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: Public data and code is in anonymous repo. |
| Open Datasets | Yes | Datasets. For evaluation, we adopt two large-scale real-world datasets with different image resolutions: CIFAR-10 (32 ร 32) and Image Net (64 ร 64), following standard protocol. ... C.1 Datasets CIFAR-10 4: This dataset contains 60,000 32 ร 32 color images... Image Net 5: Image Net is one of the most influential benchmarks in computer vision. |
| Dataset Splits | Yes | For CIFAR-10: This dataset contains 60,000 32 ร 32 color images evenly distributed over 10 distinct classes (6,000 images per class). It is split into a training set of 50,000 images and a test set of 10,000 images... For Image Net: It comprises over 1.2 million training images and around 50,000 validation images, categorized into 1,000 diverse classes. |
| Hardware Specification | Yes | Time Efficiency. We measure the training throughput of the trajectory generator gฯ( ) under different combinations of loss functions to evaluate time efficiency. Throughput results, reported in images per second per GPU, are summarized in Table 11. ... Table 11: Time efficiency comparison (imgs/sec. on H100 GPU) of SCo T under different loss functions across datasets. |
| Software Dependencies | No | We compute the partial derivative Gฯ(xt, t, s)/ โs using PyTorchโs torch.autograd.grad function... We compute FLOPs using the calflops utility... We adopt clean-fid for FID computation... We follow the same evaluation protocol as clean-fid... Our implementation follows the version in the ADM repository. |
| Experiment Setup | Yes | Training and Sampling Hyperparameters. To ensure stable convergence and fair comparison during the training of SCo T, we adopt a higher learning rate for smaller datasets (e.g., CIFAR-10) and scale the batch size appropriately for larger datasets (e.g., Image Net). We use the Adam optimizer for both settings, and enable mixed-precision training (FP16)... A complete list of hyperparameter settings is provided in Table 8. ... Table 8: Training configuration of SCo T in different model sizes on CIFAR-10 and Image Net. |