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..
Surface-Aware Feed-Forward Quadratic Gaussian for Frame Interpolation with Large Motion
Authors: Zaoming Yan, Yaomin Huang, Pengcheng Lei, Qizhou Chen, Guixu Zhang, Faming Fang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments Metrics. We use common quantitative metrics: Peak Signal-To-Noise Ratio (PSNR) and Structural Similarity Image Metric (SSIM)... Datasets. For fair comparison, we follow the training and testing datasets established by the large motion benchmark [12]. ... 5.1 Comparison with Previous Methods ... 5.2 Ablation Study |
| Researcher Affiliation | Academia | 1 School of Computer Science and Technology, East China Normal University, Shanghai, China. |
| Pseudocode | No | The paper describes the methodology using mathematical formulations (e.g., Eq. 7, 8, 9, 10, 11, 12, 13, 14, 15) and textual descriptions of modules and their interactions, but does not present a distinct pseudocode or algorithm block. |
| Open Source Code | No | The Neur IPS Paper Checklist asks: '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?' The answer provided by the authors is '[No]'. |
| Open Datasets | Yes | Datasets. For fair comparison, we follow the training and testing datasets established by the large motion benchmark [12]. For training, we follow the setting [12], utilizing both the Vimeo90K and X4K1000FPS (X-Train) datasets. Vimeo90K [13] consists of 51,312 triplets with a resolution of 448 × 256, featuring an average motion magnitude between 1 and 8 pixels. X4K1000FPS (X-Train) [17] contains 4,408 clips at a resolution of 768 × 768, with each clip comprising 65 consecutive frames. We evaluate its performance following the large motion benchmark introduced by SGM-VFI [12]. X-Test-L [17, 12] with the largest temporal gap, as our primary benchmark for evaluating large motion scenarios. We also choose the 0th and 32nd frames as input and evaluate the quality of the synthesized 16th output frame. SNU-FILM-L [80, 12] is the most challenging half of the SNU-FILM hard and extreme, with 155 triplets each. Xiph-L [12] is constructed based on the original Xiph dataset [83] by doubling the input temporal intervals and retaining the most challenging half of the data to form this benchmark. |
| Dataset Splits | Yes | For training, we follow the setting [12], utilizing both the Vimeo90K and X4K1000FPS (X-Train) datasets. ... We evaluate its performance following the large motion benchmark introduced by SGM-VFI [12]. X-Test-L [17, 12] with the largest temporal gap, as our primary benchmark for evaluating large motion scenarios. We also choose the 0th and 32nd frames as input and evaluate the quality of the synthesized 16th output frame. SNU-FILM-L [80, 12] is the most challenging half of the SNU-FILM hard and extreme, with 155 triplets each. Xiph-L [12] is constructed based on the original Xiph dataset [83] by doubling the input temporal intervals and retaining the most challenging half of the data to form this benchmark. |
| Hardware Specification | No | The paper mentions training details such as optimizer, learning rate, batch size, and epochs, but it does not provide specific hardware details like GPU models, CPU types, or memory used for the experiments. |
| Software Dependencies | No | We optimize the loss using Adam in Py Torch framework." The paper only mentions 'PyTorch framework' without specifying a version number or other software dependencies with their versions. |
| Experiment Setup | Yes | Implementation Detail. We optimize the loss using Adam in Py Torch framework. The cosine scheduler schedules the learning rate from 1e 4 to 1e 6. Standard data augmentation techniques, such as flipping, rotation, and cropping, are applied to the data with a size of 518 × 280. We train our model on the training datasets with a batch size 16 for 800 epochs. |