This is a playground for experimenting with GANs. It is a collection of scripts that make it easy to train GANs for image generation via TensorFlow.

After about 40 min of training (50 epochs // RTX3080) on the CUHK CompCars Dataset, a GAN was be able to generate images that look like this:

While these are obviously not real cars, you would most likely agree that some of these images are showing cars or something similiar to cars. With additional training or adjustments to the model architecture, it should be performing better. You can adjust the model architecture and hyperparameters in the models.py and configuration.py files.

To install all relevant libraries, run the following command:

pip install -r requirements.txt

Once you are ready to go, create a raw_data folder at project root level and put the images in there. You can have them placed in subfolders as well and the dataset generator will automatically find them.

To train a model you can edit configuration.py for your desired settings or leave them as is and run the main script:

python main.py

A number of images will be generated within each epoch and saved to the image_output folder. After the completion of a training, both the generator and the discriminator will be saved to the saved_models folder. You can then use the saved models to generate new images after loading them with tf.keras.models.load_model.