Approaches to Generating Pokemon Sprites
In this project I will document various approaches to generating never before seen Pokemon sprites. The current best model will be displayed below but I will document all of the approaches to date in the sections below.
Current Best Model
The image can take a while depending on OS and internet speed.
This current model has to be significantly compressed to share it over the internet. Additionally, due to the compression the model must run on the CPU and cannot utilize the GPU with OpenGL. The model is implemented in both TensorFlow JS and ONNX Runtime Web for maximum compatibility.
The ONNX Runtime Web does not work for iOS due to a WebAssembly bug; please use the TFJS option.
Generating additional sprites quickly may crash the page if the OS garbage collection is not performed quickly (this is mostly a problem on mobile devices).