With a standard library running a model, some of what happens is out of your hands. With my own inference engines I control the whole pipeline, and I can see exactly where randomness is added at each denoising step.
See also
- Synthesizing - long-form music with MusicGen, where the same seed didn't always give the same music