Learning Modulated Transformation in GANs

Authors: Ceyuan Yang, Qihang Zhang, Yinghao Xu, Jiapeng Zhu, Yujun Shen, Bo Dai

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments suggest that our approach can be faithfully generalized to various generative tasks, including image generation, 3D-aware image synthesis, and video generation, and get compatible with state-of-the-art frameworks without any hyper-parameter tuning. It is noteworthy that, towards human generation on the challenging Tai Chi dataset, we improve the FID of Style GAN3 from 21.36 to 13.60, demonstrating the efficacy of learning modulated geometry transformation.
Researcher Affiliation Collaboration Ceyuan Yang1 Qihang Zhang2 Yinghao Xu2 Jiapeng Zhu3 Yujun Shen4 Bo Dai1 1Shanghai AI Laboratory 2CUHK 3HKUST 4Ant Group
Pseudocode No The paper provides mathematical equations and descriptions of the method, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Code and models are available at https://github.com/limbo0000/mtm.
Open Datasets Yes We employ several challenging benchmarks to evaluate the efficacy of our module. Firstly, we use Image Net [10]... Additionally, we utilize LSUN Church[58] and Cat [58]... Furthermore, we employ Tai Chi [41]... Regarding video generation, we use Sky Timelapse [53]... and Tai Chi [41]... Moreover, we use You Tube Driving[60]...
Dataset Splits No The paper lists various datasets used for evaluation but does not specify explicit training, validation, and test dataset splits (e.g., percentages, sample counts, or predefined split IDs) for its experiments.
Hardware Specification Yes Although we train each model on a server with 8 A100 GPUs, it requires no large GPU memory and allows other researchers to reproduce easily.
Software Dependencies No The paper mentions relying on 'official implementation of Style GAN3 and Style GAN-V' and 'official implementation of EG3D', but it does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We keep all training hyper-parameters (e.g., batch size, learning rates, number of iterations, and coefficients of R1 penalty) identical for both baseline generators and the improved ones (i.e., w/ MTM).