hajar001

Generate Faces using the app : [https://stylegan-face-generation.streamlit.app/] This model has been pushed to the Hub using the PytorchModelHubMixin integration: - Code: https://github.com/HajarHAMDOUCH01/Face-Generator-StyleGAN-PyTorch - Paper: [https://arxiv.org/abs/1812.04948] - Paper: [https://arxiv.org/abs/1912.04958] - Dataset(128128 , 65k from FFHQ Dataset): [https://www.kaggle.com/datasets/hajarhamdouch/ffhq-128128-65k] This model generates realistic human faces at 128×128 resolution using a StyleGAN1 architecture trained with StyleGAN2 regularization techniques. - Architecture: StyleGAN - Training: Lazy R1 regularization + Path Length Regularization - Dataset: FFHQ (Flickr-Faces-HQ) - Resolution: 128×128 pixels - Latent dimension: 512 - Features: Style mixing, noise injection, truncation trick The `truncationpsi` parameter controls the trade-off between quality and diversity: - `1.0`: Maximum diversity, lower quality - `0.7`: Balanced (recommended) - `0.5`: Higher quality, less diversity - Epochs: 38+ (ongoing training) - Optimizer: Adam (β₁=0.0, β₂=0.99) - Learning Rates: G=0.0002, D=0.0002 - Regularization: - R1 gamma: 5.0 (applied every 16 iterations) - Path Length Regularization: 1.0 (applied every 16 iterations) - Batch Size: 32 - Style Mixing Probability: 0.8 This is a hybrid approach: StyleGAN1 architecture (AdaIN-based style modulation) trained with StyleGAN2 regularization techniques (R1 + PLR). This combines the simplicity of StyleGAN1 with the improved training stability of StyleGAN2. If you use this model, please cite the original StyleGAN papers: - Code Repository: https://github.com/HajarHAMDOUCH01/Face-Generator-StyleGAN-PyTorch - StyleGAN Paper: https://arxiv.org/abs/1812.04948 - StyleGAN2 Paper: https://arxiv.org/abs/1912.04958 - FFHQ Dataset: https://github.com/NVlabs/ffhq-dataset