Depth Anything V4 Uses Riemannian Flow Matching for Dynamic 4D Scene Reconstruction
RexDouglass · x · 2026-08-21
Depth Anything V4 (DAV4) introduces a framework for dynamic 4D scene reconstruction from monocular video. Its key innovation is applying Riemannian Flow Matching (RFM) to 4D Gaussian Splatting parameters, defining probability paths directly on non-Euclidean manifolds for scale, rotation, and opacity.
Key Improvements & Data:
- RFM improves the F-score to 0.806, compared to 0.762 for a deterministic MLP baseline with identical architecture and optimization (a gain of +0.044).
- Reduces invalid Gaussians from 12.3% to 0.01%.
- Outperforms prior Depth Anything models and per-scene 4D-GS on dynamic reconstruction and novel-view synthesis.
- Requires no human-annotated depth labels for training.
- Pre-training costs 360 GPU-hours, amortizable for large-scale deployment.
DAV4 represents a strategic shift from frame-wise depth prediction to probabilistic 4D scene understanding.
More from Multimodal
- fal launches Worlds for real-time creation and interaction with AI-generated worlds — Scobleizer · 2026-10-16
- Gummy worms as a physics test: Gemini 4 Argon vs. Claude Opus 5.5 elasticity sims — DynamicWebPaige · 2026-10-16
- 260M-parameter looped diffusion transformer beats 6x larger baselines with 5x less compute — burkov · 2026-10-16
- Seedance 2.5 demo: stretching one second of an explosion into a 30-second shot — LudovicCreator · 2026-10-16
- Octopus Short Film Made with Grok Imagine, Upscaled via Topaz — creatoroff · 2026-10-16
- Open-sourced 61-second computing history film generated entirely by Muse Spark — sophiamyang · 2026-10-16