STARFlow-V: End-to-End Video Generation Modeling Based on Normalizing Flows

STARFlow-V is an end-to-end model that applies normalizing flows (NF) to video generation. While naturally providing end-to-end training and likelihood estimation, it presents a new alternative to the diffusion-dominated video generation landscape.

Building on the recently proposed STARFlow, the model uses a global-local architecture that operates in a spatiotemporal latent space. It restricts causal dependencies to the global latent space alone while preserving local interactions within frames, reducing error accumulation over time.

On top of this, flow-score matching is added to improve the consistency of autoregressive generation with a lightweight causal denoiser, and video-aware Jacobi iteration reformulates internal updates into parallelizable iterations to boost sampling efficiency.

• The same model supports text-to-video, image-to-video, and video-to-video.
• In experiments, it showed strong visual quality and temporal consistency compared to diffusion-based baselines, along with practical sampling throughput.
• The authors present this as the first evidence that normalizing flows are capable of high-quality autoregressive video generation, positioning it as a promising direction toward world models.