Google DeepMind's D4RT Wins CVPR Award for Dynamic 4D Scene Reconstruction

Google

Jun 8, 2026 · Updated Jun 20, 2026

Google Research and Google DeepMind's paper on D4RT, a model for dynamic 4D scene reconstruction, received the CVPR 2026 Best Paper Award. This recognition highlights a new feedforward approach that efficiently reconstructs complex geometry and motion from video, unifying multiple tasks with improved speed and accuracy.

The paper "Efficiently Reconstructing Dynamic Scenes One D4RT at a Time" introduces D4RT, a feedforward model for reconstructing dynamic scenes from a single video. It uses a unified transformer architecture to jointly infer depth, spatio-temporal correspondence, and camera parameters. A novel querying mechanism avoids heavy computation of dense, per-frame decoding.

Award: CVPR 2026 Best Paper Award
Model Architecture: Unified transformer
Core Innovation: Novel querying mechanism
Efficiency: 18-300x faster than prior methods
Performance: New state of the art in Dynamic 4D Reconstruction and Tracking
Key Capabilities: Point Track, Point Cloud, Depth Map, Extrinsics, Intrinsics

D4RT addresses the challenge of reconstructing dynamic scenes, outperforming prior methods in speed and accuracy. It offers a lightweight, scalable solution by avoiding complex multiple decoders and expensive test-time optimization. This capability is critical for real-time understanding of changing environments.

D4RT provides a unified framework for 4D perception, enabling point cloud reconstruction, depth map estimation, and 3D point tracking. Its efficient, query-based decoder allows independent prediction of any point's 3D position in space and time, suitable for dynamic environments. Google Research previously introduced the D4RT model as a unified solution for 4D scene reconstruction and tracking Google Research Introduces D4RT for Unified 4D Scene Reconstruction.

View the full update on arxiv.org

Google Research

@GoogleResearchJun 7

Congratulations to the authors of "Efficiently Reconstructing Dynamic Scenes One D4RT at a Time", recipient of the #CVPR2026 Best Paper Award! 🏆 https://t.co/iATvKX94aF @GoogleDeepMind https://t.co/Ot4ZDJL8SK

24392

View on X

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

See all AI news & updates from Google →

Keep reading

Google Research Introduces D4RT for Unified 4D Scene Reconstruction

Google Research is introducing D4RT, a unified AI model that reconstructs and tracks dynamic scenes across space and time from a single video. This model advances computer vision by efficiently inferring depth, spatio-temporal correspondence, and camera parameters, setting a new state of the art for 4D scene understanding.

NVIDIAJun 6

NVIDIA PixelDiT Achieves State-of-the-Art Pixel-Space Image Generation by Removing Autoencoders

NVIDIA Research's PixelDiT, a single-stage image generation model, was selected as a best paper finalist at CVPR 2026. It removes the autoencoder step common in most image generation models, learning the diffusion process directly in pixel space to preserve fine details and achieve state-of-the-art performance among pixel-space generative models. This approach addresses a key limitation in image quality by eliminating a lossy compression step.

Google DeepMindMay 22

Google Integrates Street View Into Project Genie for Real World Simulations

Google DeepMind's Project Genie now allows users to transform real-world U.S. locations from Google Maps Street View into interactive, navigable 3D environments. By anchoring generative world models in real-world imagery, the update shifts AI simulation from purely imaginative landscapes to playable versions of actual places.

RunwayMar 20

Runway Unveils Real-Time Video Model Built with NVIDIA Hardware

Runway shared a research preview of a real-time video generation model developed with NVIDIA, running on Vera Rubin hardware. HD video generates instantly — time-to-first-frame under 100ms — opening a fundamentally different design space for video generation and world simulation.