In this project I train an unconditional DDPM to sample images of Pokemon using the following dataset https://huggingface.co/datasets/huggan/pokemon. The method used is similar to that of Ho et al. (2020) [https://arxiv. org/abs/2006.11239] with the changes to the scheduler and loss described in Nichol et al. (2021) [https://arxiv.org/abs/2102.09672]. The network is a simple UNet with self attention applied at the 32x32 and 16x16 layers and is used to predict the added noise.
To use the model run the following script:
python src/sample.pyThis should produce images similar to the following:
The model seems to be doing an ok job at generating new Pokemon, but most of them resemble colourful forms of Muk. This is due to the relatively small dataset size.