(Sorry if this was already suggested. I tried searching the github issues and discord and couldnt find anything related)
The newest video on the main channel gave me the idea:
One could make detailed Gaussian Splats, remove the "background" and programmatically render photo realistic images from different angles (even with different lighting etc.) to use as data for training?