WIP Project

Pokemon Visual Transformer Diffusion

This is a transformer based model trained for denoising task on pokémon sprites dataset from first generation. The goal is to produce new and original sprites while keeping a coherent style. This is still a work in progress project.

Best results for now

How tu use it

git clone https://github.com/MattiasKockum/PokemonAIGen.git
cd PokemonAIGen
python -m venv venv
source venv/bin/activate
#sudo mount -o remount,size=16G /tmp # This might be needed
pip install -r requirements.txt

Fill a .env file with your own data

role = "..." # Get it from AWS
pt_mnist_model_data = "..." # You get it by running launch_training.py
wandb_api_key = "..." # Get it from Weights And Biases

python prepare_data.py
python launch_training.py

python deploy.py

Look into outputs directory

TODO

Normalize test and train loss

Add sliding into data augmentation

Make the noise always the same on testing !

Early stopping

Regularization

Add color (multiple channels)

Data augmentation

Here are exemples of data augmentation done to ensure better robustness of the model.