Denoising Diffusion Probabilistic Doom

80 epochs of training
100 epochs of training

Inspiration

Recent advances in AI generated images such as Dalle-2 and Stable Diffusion

What it does

Uses machine learning to look at all the wall textures present in the game Doom and then predicts what new textures would look like.

How we built it

Texture files were dumped from the Doom 2 WAD file, loaded into python for image augmentation to create a dataset of over 4000 images. A denoising diffusion probabilistic model (DDPM) was then trained on these images.

Challenges we ran into

Since Doom textures are 128 x 128 resolution, which is significantly larger than most toy image generation datasets, training of the model was incredibly slow so I had to rent out cloud GPU power and adapt the training scripts for multi-gpu parallel processing.

Accomplishments that we're proud of

The fact that the images don't look like total garbage

What we learned

How to train DDPM AIs

What's next for Denoising Diffusion Probabilistic Doom

Doom wall textures are meant to be able to tile seamlessly horizontally, however the model has no intrinsic way of knowing this and is difficult to infer from the dataset so in the future I would like to implement a loss function that incorporates tileability by looking at the delta between the left and right side of the generated images