🎨 Generate traditional-style Pokémon cards using Deep Convolutional Generative Adversarial Networks (DCGANs).
2024-11-24.12-57-30.mp4
This project utilizes a DCGAN to create realistic Pokémon cards, focusing on traditional designs from the Pokémon Trading Card Game. The model generates visually convincing cards by training on a dataset of over 11,000 card images spanning all generations, excluding full-art types like EX and GX to maintain a focus on the classic card format.
- Generate unique, traditional-style Pokémon card designs.
- Trained on a dataset of over 11,000+ Pokémon cards with consistent rectangular dimensions (600x825 pixels).
- Built with PyTorch and trained on an NVIDIA GeForce RTX 3050 (4GB).
- Source: Pokemon TCG Dataset
- Size: 11,044 preprocessed images
- Preprocessing Steps:
- Verified image validity and removed invalid or incompatible files.
- Resized images to consistent dimensions of 600x825 pixels.
- Converted images to RGB mode for uniformity.
- Takes a 100-dimensional latent vector (
z) and outputs a full-color image (96x64 pixels). - Architecture:
- 5 transposed convolutional layers.
- Batch normalization for stability.
- Final activation:
Tanh.
- Classifies input images (96x64 pixels) as real or fake.
- Architecture:
- 5 convolutional layers.
- Leaky ReLU activations.
- Final activation:
Sigmoid.
Both models use custom weight initialization based on the DCGAN paper.
-
Loss Graphs: Generator and Discriminator losses during training:
-
Turing Test: The Real Vs Fake:
- Rectangular Card Dimensions: Adjusted the dataset and architecture to handle non-square images.
- Clone this repository:
git clone https://github.com/olivergrabner/pokemon-card-dcgan.git