Research map / Video / Video generation
Scene Splatter: Momentum 3D Scene Generation from Single Image with Video Diffusion Model
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.
Scene Splatter iteratively generates a 3D scene from a single image by using latent-level and pixel-level "momentum" from rendered Gaussian splats to guide a video diffusion model (ViewCrafter) and refine 3DGS.
Problem
Generating consistent, high-fidelity 3D scenes from a single image with video diffusion models is hard because generated videos are short and change existing scene content, causing artifacts when reconstructing 3D.
Approach
Construct noisy samples from rendered latent features as latent-level momentum to keep existing content consistent, then blend with a pixel-level momentum (scale-map-weighted) so unseen regions are still generated freely; iterate render-generate-finetune of Gaussians to build scenes beyond video length limits.
Key results
- On the RealEstate10K easy set, Scene Splatter reaches 20.95 PSNR / 0.800 SSIM / 0.145 LPIPS vs ViewCrafter's 20.70 / 0.794 / 0.159.
- On the hard set, Scene Splatter reaches 17.62 PSNR vs 15.63 for ViewCrafter and 14.41 for Flash3D.
- Ablation: removing pixel-level momentum drops PSNR from 19.81 to 16.31; removing latent-level momentum drops it to 18.52.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Scene Splatter's cascaded momentum instead of ViewCrafter | Scene Splatter · RealEstate10K easy set | PSNR: 20.70 -> 20.95 | Helped |
| use Scene Splatter's cascaded momentum instead of ViewCrafter | Scene Splatter · RealEstate10K hard set | PSNR: 15.63 -> 17.62 | Helped |
| add video diffusion model (ablation) | Scene Splatter · Rendering results | PSNR: 16.14 -> 19.81 | Helped |
| add latent-level momentum (ablation) | Scene Splatter · Rendering results | PSNR: 18.52 -> 19.81 | Helped |
| add pixel-level momentum (ablation) | Scene Splatter · Rendering results | PSNR: 16.31 -> 19.81 | Helped |
Surprising findings
- Without the video diffusion model, the framework degenerates to Flash3D and loses about 3.7 dB PSNR, while CogVideoX with no momentum scores lower than Flash3D on the easy set in PSNR.
Limitations
- Slower than regression-based methods due to video diffusion and iterative optimization.
- Restricted to static scenes; cannot recover 4D scenes.
- Evaluation limited to a RealEstate10K subset with rendering metrics only.
Experimental setup
Datasets: RealEstate10K
Benchmarks: RealEstate10K subset (easy and hard sets)
Metrics: PSNR SSIM LPIPS
Baselines: Flash3D CogVideoX ViewCrafter
Builds on: 3D Gaussian Splatting ViewCrafter Flash3D latent video diffusion
Closest papers by meaning
- Wonderland: Navigating 3D Scenes from a Single Image (CVPR 2025 · similarity 0.91)
- Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs (CVPR 2025 · similarity 0.89)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.89)
- 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors (NeurIPS 2024 · similarity 0.89)
- BloomScene: Lightweight Structured 3D Gaussian Splatting for Crossmodal Scene Generation (AAAI 2025 · similarity 0.89)
- Splatter Image: Ultra-Fast Single-View 3D Reconstruction (CVPR 2024 · similarity 0.89)
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.