STARFlow-V

TL;DR

Abstract

Normalizing flows (NFs) are end-to-end likelihood-based generative models for continuous data, and have recently regained attention with encouraging progress on image generation. Yet in the video generation domain, where spatiotemporal complexity and computational cost are substantially higher, state-of-the-art systems almost exclusively rely on diffusion-based models. In this work, we revisit this design space by presenting STARFlow-V, a normalizing flow-based video generator with substantial benefits such as end-to-end learning, robust causal prediction, and native likelihood estimation. Building upon the recently proposed STARFlow, STARFlow-V operates in the spatiotemporal latent space with a global-local architecture which restricts causal dependencies to a global latent space while preserving rich local within-frame interactions. This eases error accumulation over time, a common pitfall of standard autoregressive diffusion model generation. Additionally, we propose flow-score matching, which equips the model with a light-weight causal denoiser to improve the video generation consistency in an autoregressive fashion. To improve the sampling efficiency, STARFlow-V employs a video-aware Jacobi iteration scheme that recasts inner updates as parallelizable iterations without breaking causality. Thanks to the invertible structure, the same model can natively support text-to-video, image-to-video as well as video-to-video generation tasks. Empirically, STARFlow-V achieves strong visual fidelity and temporal consistency with practical sampling throughput relative to diffusion-based baselines. These results present the first evidence, to our knowledge, that NFs are capable of high-quality autoregressive video generation, establishing them as a promising research direction for building world models.

Method Pipeline

Figure: STARFlow-V pipeline. The model processes text prompts and noise through a Deep Autoregressive Block (global temporal reasoning) to produce intermediate latents, which are then refined by Shallow Flow Blocks (local within-frame details). A Learnable Causal Denoiser (trained via Flow-Score Matching) cleans the output. The model is trained end-to-end with two objectives: Maximum Likelihood for the flow and Flow-Score Matching for the denoiser.

Key Contributions

Model Details

STARFlow-V is trained on 70M text-video pairs and 400M text-image pairs, with a final 7B parameter model that can generate 480p video at 16fps. The model operates in a compressed latent space and leverages the invertible nature of normalizing flows to natively support multiple generation tasks without any architectural changes or retraining.

Explore the Results

Navigate through the tabs above to see our model's capabilities across different generation tasks. Each category demonstrates specific aspects of STARFlow-V, from standard text-to-video generation to long-form video creation and comparisons with diffusion-based baselines.

BibTeX

If you find STARFlow-V useful in your research, please consider citing our work:

@article{gu2025starflowv,
  title={STARFlow-V: End-to-End Video Generative Modeling with Scalable Normalizing Flows},
  author={Gu, Jiatao and Shen, Ying and Chen, Tianrong and Dinh, Laurent and Wang, Yuyang and Bautista, Miguel \'Angel and Berthelot, David and Susskind, Josh and Zhai, Shuangfei},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2025}
}