Recover the dynamic scene
Build a 4D representation with aggregated static geometry and time-indexed moving objects, then render a coarse view along the target trajectory.
DrivENVS
Generative Novel View Synthesis from Monocular Driving Videos
1 University of Maryland, College Park
2 Stack AV
From a single driving video, DrivENVS reconstructs dynamic scenes and synthesizes coherent video from new camera trajectories using geometry-aware, few-step image diffusion.
Watch intro video ↓
Not a new world.
A new way to see it.
DrivENVS grounds each novel camera view in motion-aware 4D geometry, retrieves relevant observations from the source video, and refines coarse renderings into temporally coherent output.
Look back at relevant frames
Select references using 3D co-visibility; combine top-1 joint attention, multi-reference cross-attention and relative camera-time encoding.
Refine, frame by frame
Apply few-step image diffusion using prior generated frames to maintain temporal consistency over long sequences.
The camera moves.
The scene persists.
Explore camera translations. Source and target cameras share the same orientation; only the target position changes. Drag the scene to inspect the geometry.
Translation direction
Camera displacement
Loading interactive 3D scene…
Source camera
Target camera
Drag to rotate · Scroll to zoom
Qualitative Results.
Explore synchronized input and synthesized video pairs across left, right and upward camera shifts, with 1-meter and 2-meter displacements.
Quality without
the video-DiT cost.
Our evaluation spans nuScenes, Waymo and Argoverse 2, including translated and side-facing target views.
28.8%
Average FVD reduction against the related
image-diffusion baselines in the paper.
20 GB
Approximate peak inference memory,
compared with 88 GB for NeoVerse
in the reported comparison.
4 steps
Few-step rectified-flow refinement
for each generated frame.
