Its encoder maps an input, such as an image, into parameters of a probability distribution in a lower-dimensional latent space. A decoder then reconstructs the input or creates a new sample from a latent value.
Training balances reconstruction quality with a regularization term that makes the learned latent space suitable for sampling. VAEs are used for representation learning, anomaly detection and controlled generation, although their outputs can be less sharp than those of some other image generators.
VAEs appear in model-collapse research because repeated fitting on generated samples can reveal how a learned distribution changes across generations. Results depend on the model family and assumptions, rather than proving that all machine-generated training data are damaging.
Compare reconstruction quality and sample diversity, and measure performance against untouched real examples. When interpreting recursive-training studies, check whether samples replaced or supplemented the original data.