Apple Unveils STARFlow-V: A Normalizing Flow Breakthrough Challenging Diffusion Models in Video Generation

Apple Machine Learning Research has announced STARFlow-V, an innovative normalizing flow-based model poised to redefine end-to-end video generative modeling, with a paper slated for publication at CVPR in April 2026. Normalizing flows (NFs), which are end-to-end likelihood — based generative models for continuous data, have recently seen encouraging progress in image generation. With STARFlow-V, researchers revisit this design space for video generation, distinguishing the model through a sophisticated global — local architecture operating within a spatiotemporal latent space.

The introduction of STARFlow-V marks a pivotal moment for normalizing flows in the video generation domain, an area where diffusion models have almost exclusively held the state — of-the-art position due to the inherent spatiotemporal complexity and computational demands. STARFlow-V offers substantial benefits, including end-to-end learning, robust causal prediction, and native likelihood estimation. This development directly challenges the prevalent reliance on diffusion models by demonstrating a viable and effective alternative for generating dynamic visual content.

Building upon the recently proposed STARFlow, an architecture known for scaling latent normalizing flows for high-resolution image synthesis, STARFlow-V introduces further innovations specific to video. Central to its advanced capabilities is a novel technique called flow-score matching, which equips the model with a lightweight causal denoiser. This denoiser significantly improves video generation consistency in an autoregressive fashion, ensuring smoother and more coherent temporal sequences. Furthermore, to enhance sampling efficiency, STARFlow-V employs a video — aware Jacobi iteration scheme. This scheme cleverly recasts inner updates as parallelizable iterations, allowing for faster processing without compromising the crucial aspect of causality within the generated video.

A key advantage stemming from STARFlow-V's invertible structure is its native support for a wide array of generation tasks. The same model can seamlessly handle text-to-video, image — to-video, and video — to-video generation, providing remarkable versatility for various creative and analytical applications. Empirically, STARFlow-V has demonstrated strong visual fidelity and temporal consistency, achieving practical sampling throughput when benchmarked against diffusion — based baselines. These results highlight its potential for real-world application, offering a compelling performance alternative in a computationally intensive field.

The empirical success of STARFlow-V provides what the researchers describe as the first substantial evidence that normalizing flows are indeed capable of high-quality autoregressive video generation. This achievement establishes normalizing flows as a promising direction for fundamental research, particularly in the ambitious pursuit of building world models capable of understanding and simulating complex dynamic environments. The work represents a significant step forward in generative AI, potentially broadening the architectural landscape for future developments in video synthesis.

This groundbreaking research was led by a team of distinguished scientists including Jiatao Gu from the University of Pennsylvania, Ying Shen from the University of Illinois Urbana — Champaign, and Apple researchers Tianrong Chen, Laurent Dinh, Yuyang Wang, Miguel Ángel Bautista, David Berthelot, Josh Susskind, and Shuangfei Zhai. Their collective expertise has culminated in a model that not only pushes the boundaries of video generation but also re-evaluates the efficacy of normalizing flows in complex spatiotemporal domains, building upon earlier work such as STARFlow, which was presented at NeurIPS on June 30, 2025.