Reviving the Past: Apple Research Harnesses a Forgotten AI Technique for Image Generation

Apple’s recent research introduces a new approach to generative image models through Normalizing Flows (NFs), proposing that they could be a viable alternative to current methods like diffusion and autoregressive models. NFs learn to convert real-world data into structured noise and back, allowing for precise likelihood calculations, which is beneficial for probability-sensitive tasks. Apple’s first study presents TarFlow, which employs Transformer blocks for autoregressive image generation by predicting image patches in sequence, bypassing the tokenization process that often compromises image quality. The second study introduces STARFlow, which operates in latent space to generate compressed image structures before enhancing resolution with a decoder, thus improving scalability for high-resolution images. Unlike OpenAI’s GPT-4o, which generates images token-by-token, Apple focuses on a model that aims to efficiently manage resources for mobile devices. Both companies are innovating beyond diffusion models, with Apple emphasizing on-device capability and efficiency.

Source link