This repo contains a Generative Adversarial Network (GAN) model trained using StyleGAN-ada-2 to generate RPG (Role-Playing Game) sprites. Each sprite consists of four images showing it from different perspectives: front, behind, right, and left. The model was trained on a dataset of 912 sprites styled after old RPG games.
- Architecture: StyleGAN-ada-2
- Training Steps: Almost 300 epochs
- FID50k_full Score: 8.2311772
- Link to the model
- The training set contains 912 RPG sprites mostly from old games.
- Each image has a shape of
[3, 128, 128]and are converted to RGB with a resolution of 128x128 pixels. - Each image consists of 4 aspects of the character(front, rear, right, left).
Judging by the fid50k_full score, The model has pretty much captured the gist of RPG art style, producing sprites that exhibit the commonly known characteristics of RPG game sprites.
To generate RPG sprites using the trained model, follow these steps:
- Clone the repository.
- Clone the official repo of StyleGAN2-ada
- Install the necessary dependencies(you might need to debug the
conv2d_gradfix.pyandgrid_sample_gradfix.pyyourself to avoid PyTorch version mismatch errors) - Run the provided scripts or integrate the model into your project for sprite generation.
Real Sprites:
Synthesized Sprites:
