This paper shows that the best VAEs for image generation are the ones whose latents neatly separate object attributes, a property called semantic disentanglement.
The paper turns one flat picture into a neat stack of see‑through layers, so you can edit one thing without messing up the rest.