The repository is an implementation, based on tensorflow, of a diffusion model, which has been trained to generate images of faces resembling characters from the famous series "The Simpsons". Sampling (and so the training) is completely unconditional, letting the model produce whatever it comes to.
Following results have been sampled after training the model for 100 epochs (more training details in the config.py file). Weights can be found under the weights folder, and they are automatically loaded when synthesizing new images. The sampling procedure is the standard one: starting from a random gaussian noise, the image is gradually denoised, thanks to the reverse diffusion process, finally being transformed into a distribution similar to the ones encountered during the training phase (in this case a Simpsons-like face).
Despite the low resolution of 64x64 pixels (due to hardware's limitations), these results are noticeably good; especially if compared to similar generations coming from various GAN-based models. Many images reproduce closely the most relevant characters, such as the whole Simpsons family. Others create new faces, with some of them clearly reflecting a (good) combination of facial treats (or typical expressions) of different characters.
The model is capable of sampling infinitely many unique (and sometimes curious) faces, with a fixed resolution of 64x64 pixels. In order to synthesize new faces, run the generator.py script:
python generator.py "save_folder"
There is only one optional argument:
--n: set the number of images to sample (to be chosen accordingly to GPU's memory, default=8)
Results are saved into the specified folder as "sample i.jpg", where i ranges from 1 to the requested number of images.
Training the model from scratch is as simple as editing the hyperparameters inside the config file. The preset that can be currently found is the one that has been adopted for the purpose of this work. Though, the model can be re-trained on whichever dataset, that should be placed inside data/training dataset.
The network has been trained using the dataset named "Simpsons Faces", available on Kaggle. This one can be used to experiment further, trying to reproduce these results or to improve on them (e.g., by increasing the resolution, increasing the number of attention heads, introducing a preprocessing pipeline etc.).
To start a new training session (or recover from the last training step), run the command:
python train.py
Training parameters can be adjusted by accessing the file src/model/config.py.