SUV: Scene Future as Video Generation for End-to-End Driving
On August 4, 2026, a paper on SUV appeared on arXiv — an end-to-end driving framework that frames understanding of the future scene as a video generation task. Instead of specialized predictive modules, the authors use a pretrained video model to directly generate four video streams: appearance, semantics, relative depth, and object dynamics. The trajectory is built through joint attention to the latent representations of all these streams. On the NAVSIM-v2 benchmark, the method with a single front camera and no trajectory candidate search achieved 91.0 EPDMS on navtest and 36.9 on navhard, and on WOD-E2E it showed RFS 7.94. More on arXiv
What it means
SUV is not a video generator in the usual sense — it is a perception and planning system for autonomous driving that internally uses a video generative model to predict the future. What it shares with the video generators in our AI Video section is only the architectural approach: the video model acts as a general predictor for multiple modalities.
None of the tools in the reference — LTX-Video, CogVideoX, or Pyramid Flow — solves the end-to-end driving task or predicts scene semantics or depth. These are text-to-video or image-to-video generators that run on consumer hardware. SUV, on the other hand, is a research framework for autonomous vehicles, and the developer does not specify hardware requirements or code availability.
The work shows that video generative models can serve as a universal component for tasks far beyond entertainment, but there is no practical overlap with video creation tools yet.
