STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation

2026-05-08

AI Summary

Existing unified multimodal generation approaches struggle to achieve genuine interleaved text-image synthesis due to structural mismatches between textual and visual generation mechanisms. This work proposes STARFlow2, a framework that unifies autoregressive normalizing flows with language model architectures under a shared causal masking scheme. By integrating the Pretzel architecture, vertically interleaved pretraining of vision-language models (VLMs), a dual-path flow design, and a unified FAE latent space, STARFlow2 enables end-to-end, KV cache-friendly interleaved generation of text and images. Experimental results demonstrate strong performance on both image generation and multimodal understanding benchmarks, confirming the viability of autoregressive flows as a unified paradigm for multimodal modeling.

Abstract

Deep generative models have advanced rapidly across text and vision, motivating unified multimodal systems that can understand, reason over, and generate interleaved text-image sequences. Most existing approaches combine autoregressive language modeling with diffusion-based image generators, inheriting a structural mismatch between causal text generation and iterative visual denoising. We observe that autoregressive normalizing flows are autoregressive Transformers--sharing the same causal mask, KV-cache mechanism, and left-to-right structure as LLMs--making them the most natural paradigm for true unified multimodal generation. We present STARFlow2, built on the Pretzel architecture that vertically interleaves a pretrained VLM stream with a TarFlow stream via residual skip connections, both operating under the same causal mask. Combined with a deep-shallow flow design and a unified FAE latent space, STARFlow2 enables cache-friendly interleaved generation where both text and visual outputs directly enter the KV-cache without re-encoding. Experiments demonstrate strong performance across image generation and multimodal understanding benchmarks, validating autoregressive flows as a viable foundation for unified multimodal modeling.

Problem

• multimodal generation
• structural mismatch
• autoregressive modeling
• diffusion models
• unified modeling

Innovation

• autoregressive normalizing flows
• unified multimodal generation
• causal Transformer
• KV-cache integration
• FAE latent space