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..
OmniZoom: A Universal Plug-and-Play Paradigm for Cross-Device Smooth Zoom Interpolation
Authors: Xiaoan Zhu, Yue Zhao, Tianyang Hu, Jiaming Guo, Yulan Zeng, Renjing Pei, Fenglong Song, Huajun Feng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a comprehensive evaluation of the 1D, 2D, and 3D-TPR indexing frameworks across four FI models and four real-world smartphone platforms, considering both the base models and those finetuned on our ZI dataset. The paper includes a dedicated "4 Experiments" section, ablation studies, and presents quantitative and qualitative results using various metrics like PSNR, SSIM, and LPIPS, demonstrating empirical validation. |
| Researcher Affiliation | Collaboration | Xiaoan Zhu Zhejiang University EMAIL Yue Zhao Huawei Noah s Ark Lab EMAIL Tianyang Hu Zhejiang University EMAIL Jiaming Guo Huawei Noah s Ark Lab EMAIL Yulan Zeng Sungkyunkwan University EMAIL Renjing Pei Huawei Noah s Ark Lab EMAIL Fenglong Song Huawei Noah s Ark Lab EMAIL Huajun Feng Zhejiang University EMAIL |
| Pseudocode | No | The paper describes mathematical modeling and algorithmic steps, such as in Section 3.1 and 3.2, but does not include a distinct block explicitly labeled "Pseudocode" or "Algorithm" with structured, code-like formatting. |
| Open Source Code | No | We provide a project page at https://omnizoom.github.io/Omni Zoom/, where visual results are available, and the code/dataset will be released soon. (From the NeurIPS checklist justification: We intend to release the code and dataset upon publication. While they are not publicly available at the time of submission, the release is scheduled and will include detailed documentation to ensure full reproducibility.) |
| Open Datasets | Yes | Each model is initially trained on the Vimeo-90K [52] septuplet dataset (91,701 sequences at 448 256 resolution) using its default hyperparameters. We then finetune them on our ZI dataset. For benchmarks with ground-truth, including Vimeo90K, Inter4K [41], and UCF101 [40], we use full-reference image quality metrics. Also, the Redmi sequences are adopted from the publicly available dataset introduced in [50]. |
| Dataset Splits | Yes | Each model is initially trained on the Vimeo-90K [52] septuplet dataset (91,701 sequences at 448 256 resolution) using its default hyperparameters. We then finetune them on our ZI dataset. For benchmarks with ground-truth, including Vimeo90K, Inter4K [41], and UCF101 [40], we use full-reference image quality metrics. These are widely recognized benchmark datasets that typically use standard train/test/validation splits. |
| Hardware Specification | No | All experiments are implemented by Py Torch [35] on hardware devices equipped with 40 GB of memory. All procedures are conducted on a workstation with 40 GB of memory, consistent with our main experimental setup. This only specifies the memory amount but not specific CPU or GPU models. |
| Software Dependencies | No | All experiments are implemented by Py Torch [35] on hardware devices equipped with 40 GB of memory. While PyTorch is mentioned, a specific version number is not provided. |
| Experiment Setup | Yes | For virtual camera generation, we adopt the Adam optimizer [24] in a two-stage scheme. We first train the Gaussian model Mw for 5k iterations using the learning rate from 3DGS [23]. Then, joint optimization is performed for 30k iterations on both Mw and the Color Net, with the learning rate decaying from 1e-3 to 1e-6 after 20k steps. Our 3D-TPR framework is integrated into RIFE [17], IFRNet [25], EMAVFI [53], and AMT [27]. Each model is initially trained on the Vimeo-90K [52] septuplet dataset (91,701 sequences at 448 256 resolution) using its default hyperparameters. We then finetune them on our ZI dataset. |