StyleNeRF: A Style-based 3D Aware Generator for High-resolution Image Synthesis

Authors: Jiatao Gu, Lingjie Liu, Peng Wang, Christian Theobalt

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Style Ne RF on various challenging datasets. Style Ne RF can synthesize photo-realistic 1024^2 images at interactive rates while achieving high multi-view consistency. None of the existing methods can achieve both characteristics. Additionally, Style Ne RF enables direct control on styles. ... We evaluate Style Ne RF on four high-resolution unstructured real datasets: FFHQ (Karras et al., 2019), Met Faces (Karras et al., 2020a), AFHQ (Choi et al., 2020) and Comp Cars (Yang et al., 2015). ... Quantitative comparison We measure the visual quality of image generation by the Frechet Inception Distance (FID, Heusel et al., 2017) and Kernal Inception Distance (KID, Bi nkowski et al., 2018) in Table 1. ... 4.4 ABLATION STUDIES
Researcher Affiliation Collaboration Meta AI Max Planck Institute for Informatics The University of Hong Kong
Pseudocode No The paper describes the architecture and methods in text and with diagrams (e.g., Figure 3) but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Code and pre-trained models are available at: https://github.com/facebookresearch/Style Ne RF.
Open Datasets Yes We evaluate Style Ne RF on four high-resolution unstructured real datasets: FFHQ (Karras et al., 2019), Met Faces (Karras et al., 2020a), AFHQ (Choi et al., 2020) and Comp Cars (Yang et al., 2015). The dataset details are described in Appendix B. ... FFHQ (https://github.com/NVlabs/ffhq-dataset) ... AFHQ (https://github.com/clovaai/stargan-v2# animal-faces-hq-dataset-afhq) ... Met Faces (https://github.com/NVlabs/metfaces-dataset) ... Comp Cars (http://mmlab.ie.cuhk.edu.hk/datasets/comp_cars/)
Dataset Splits No The paper describes a progressive training strategy based on the number of images processed (T1, T2, T3) and mentions using different datasets for training, but it does not specify explicit training, validation, and test data splits with percentages or counts for reproduction.
Hardware Specification Yes All models are trained on 8 Tesla V100 GPUs for about three days.
Software Dependencies No The paper states 'We implement our model based on the official Pytorch implementation of Style GAN2-ADA', but it does not provide specific version numbers for PyTorch or Style GAN2-ADA or any other software dependencies.
Experiment Setup Yes By default, we train 64 images per batch, and set T1 = 500k, T2 = 5000k and T3 = 25000k images, respectively. The input resolution is fixed 32^2 for all experiments. ... We reuse the same architecture and parameters of Style GAN2 for the mapping network (8 fully connected layers, 100 lower learning rate) and discriminator. In addition, both the latent and style dimensions are set to 512. ... The Fourier feature dimension is set to L = 10 (Equation (1)) for both fields.