An experiment fine-tuning gpt-2 (small 100M) on a dataset of public shadertoy shaders. The auto generated comments are perhaps more entertaining than the glsl. Scaling up the model would certainly give much better results.
Training/model code were slight modifications to huggingface's example
https://github.com/huggingface/transformers/tree/master/examples/language-modeling
Shadertoy dataset was generated using this repository
https://github.com/PWhiddy/Shadertoy-Generator
