Adding rendered pairs that carry motions a training set lacks improves a model trained on that set, even when the renders come from non-photorealistic sources such as fractal signed distance fields. Adding synthetic data with uncurated motion, or more naturally-collected images from the same source, only marginally improves a model.

Three claims in one: motion beats realism as a predictor; motion-targeted pairs help even from fractal renders; the controls do not. Chapters 3 and 4 supply those three pieces of evidence. Chapter 2 explains why generation is necessary at all.