Research map / 3D vision / Depth and stereo
Optimizing 4D Gaussians for Dynamic Scene Video from Single Landscape Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Lifts 2D single-image motion (Eulerian flow) into consistent 3D motion via a 3D Motion Optimization Module, then trains 4D Gaussians from a single landscape image to produce dynamic scene videos with novel camera views.
Problem
Generating dynamic scene videos with consistent 3D motion and novel camera views from a single landscape image, where prior methods rely on pseudo-3D layered depth images and produce inconsistent motion or artifacts.
Approach
A 3D Motion Optimization Module parametrically models 3D motion (difference of two point clouds) and optimizes it so its 2D reprojections match 2D motion maps from an Eulerian-flow single-image animation model across multi-view settings; the 3D motion initializes a 4D Gaussian model trained in two stages (spatial with multi-view images, temporal with pseudo-videos from sampled views).
Key results
- Ours reaches PSNR 20.57, SSIM 0.90, LPIPS 0.14, PIQE 7.80, outperforming 3D-Cinemagraphy (17.30 PSNR) and Make-It-4D (16.98) and DynamiCrafter (14.98).
- In a user study, Ours gets 56.25% immersion, 57.81% realism, 62.81% structural consistency, and 60.25% quality preference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ours instead of 3D-Cinemagraphy | Dynamic scene video reconstruction · Dynamic Scene Video | PSNR: 17.30 -> 20.57 | Helped |
| Use Ours instead of 3D-Cinemagraphy | Dynamic scene video reconstruction · Dynamic Scene Video | SSIM: 0.83 -> 0.90 | Helped |
| Use Ours instead of 3D-Cinemagraphy | Dynamic scene video reconstruction · Dynamic Scene Video | LPIPS: 0.17 -> 0.14 | Helped |
| Use Ours instead of Make-It-4D | Dynamic scene video reconstruction · Dynamic Scene Video | PIQE: 8.30 -> 7.80 | Helped |
| Use Ours instead of 3D-Cinemagraphy | Dynamic scene video reconstruction · Dynamic Scene Video | Immersion preference (%): 31.87 -> 56.25 | Helped |
| Use Ours instead of 3D-Cinemagraphy | Dynamic scene video reconstruction · Dynamic Scene Video | Realism preference (%): 31.87 -> 57.81 | Helped |
| Use Ours instead of 3D-Cinemagraphy | Dynamic scene video reconstruction · Dynamic Scene Video | Structural Consistency preference (%): 28.75 -> 62.81 | Helped |
| Use Ours instead of 3D-Cinemagraphy | Dynamic scene video reconstruction · Dynamic Scene Video | Quality preference (%): 30.31 -> 60.25 | Helped |
Limitations
- Motion limited to Eulerian/fluid-like motion; complex articulated or human motion not handled.
- Holes from novel-view reprojection filled only by linear interpolation; no generative inpainting in core method.
- Training time increases with longer video sequences.
Experimental setup
Datasets: landscape images from Holynski et al. dataset online global landmark images
Benchmarks: landscape dynamic scene video dataset (Holynski et al.)
Metrics: PSNR SSIM LPIPS PIQE user study (immersion, realism, structural consistency, qual
Baselines: DynamiCrafter Motion-I2V 3D-Cinemagraphy Make-It-4D
Builds on: 4D Gaussian Splatting Deformable 3D Gaussians LucidDreamer Animating Pictures with Eulerian Motion Fields ZoeDepth
Closest papers by meaning
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.89)
- Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle (CVPR 2024 · similarity 0.88)
- World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video (NeurIPS 2026 · similarity 0.88)
- Instant4D: 4D Gaussian Splatting in Minutes (NeurIPS 2025 · similarity 0.87)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
- Motion4D: Learning 3D-Consistent Motion and Semantics for 4D Scene Understanding (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.