TechPrufer
Get started
Apple
Apple logoApple · Feature trail

Unified Multimodal Generation with Autoregressive Flows

1 post · Aug 2026

What this trail covers

STARFlow2 builds upon the Pretzel architecture, vertically interleaving a frozen pretrained VLM stream with a TARFlow stream via residual skip connections. This design preserves pretrained multimodal understanding, enables high-fidelity continuous image generation, and achieves structural unification under a single causal mechanism. Combined with a deep-shallow flow design and a unified FAE latent space, STARFlow2 supports cache-friendly interleaved generation where both text and visual outputs directly enter the KV-cache without re-encoding.

All Apple trails
  1. STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal GenerationSTARFlow2 presents a unified multimodal generation framework by interleaving a frozen pretrained VLM stream with a TARFlow stream using residual skip connections. This approach preserves multimodal understanding, enables high-fidelity continuous image generation, and achieves structural unification under a single causal mechanism. It supports cache-friendly interleaved generation where both text and visual outputs enter the KV-cache without re-encoding.