GLOBER: Coherent Non-autoregressive Video Generation via GLOBal Guided Video DecodER

Authors: Mingzhen Sun, Weining Wang, Zihan Qin, Jiahui Sun, Sihan Chen, Jing Liu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results demonstrate the effectiveness and efficiency of our proposed method1, and new state-of-the-art results have been achieved on multiple benchmarks.
Researcher Affiliation Academia 1Institute of Automation, Chinese Academy of Sciences (CASIA) 2School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)
Pseudocode No The paper describes its methods in prose and mathematical equations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Our codes have been released in https://github.com/iva-mzsun/GLOBER
Open Datasets Yes Table 1 and Table 2 reports the results of our model trained on the Sky Time-lapse [36], Tai Chi-HD [37], UCF-101 [38], and Webvid-10M [39] datasets for 16-frame video generation in both unconditional and conditional settings.
Dataset Splits No The paper mentions training on specific datasets (UCF101, Tai Chi HD, and Sky Time-lapse) but does not provide specific details on how these datasets were split into training, validation, or test sets for their experiments.
Hardware Specification Yes All experiments are implemented using Py Torch [40] and conducted on 8 NVIDIA A100 GPUs, with 16-precision adopted for fast training." and "Results with * are taken from PVDM and measured with a single NVIDIA 3090ti 24GB GPU. The rest are evaluated on a single NVIDIA 3090 24GB GPU by us due to lack of 3090ti.
Software Dependencies No The paper states 'All experiments are implemented using Py Torch [40]' but does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes The video auto-encoder was trained with a batch size of 40 per GPU for 80K, 40K, and 40K steps on the UCF101, Tai Chi HD, and Sky Time-lapse datasets, respectively. The loss weight λ1 and λ2 are set as 1e-6 and 0.1, respectively.