This paper shows how to make powerful image‑generating Transformers run fast on phones without needing the cloud.
This paper shows a new way to teach an autoencoder to shape its hidden space (the 'latent space') to look like any distribution we want, not just a simple bell curve.