The Digital Press

Apple has revived and enhanced a lesser-known AI technique called Normalizing Flows to generate high-quality images more efficiently, aiming for on-device use.

Why it matters: Normalizing Flows can calculate exact image likelihoods, improving control and reliability in image generation.

The big picture: Apple’s TarFlow and STARFlow models use Transformers and latent space compression to overcome previous quality and scaling issues.

The other side: OpenAI’s GPT-4o uses token-based generation in the cloud, offering versatility but at higher computational cost and latency.

Commenters say: Readers appreciate Apple’s focus on on-device AI and efficiency, while some question how it will compete with established diffusion and autoregressive models.