Research map / 3D vision / 3D generation
Wonderland: Navigating 3D Scenes from a Single Image
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.
Wonderland generates wide-scope 3D Gaussian Splatting scenes from a single image by feeding camera-guided video diffusion latents into a feed-forward latent large reconstruction model.
Problem
Generating high-quality, wide-scope 3D scenes from a single arbitrary image without multi-view data or slow per-scene optimization, while keeping geometry consistent in occluded regions.
Approach
Operate the 3D reconstruction directly in the compact latent space of a camera-controlled video diffusion model: a dual-branch (ControlNet plus LoRA) camera conditioning makes the video latents 3D-aware, and a latent large reconstruction model regresses pixel-aligned 3DGS from those latents in a feed-forward pass, trained progressively.
Key results
- On RealEstate10K, the camera-guided video model reaches FVD 153.48 and Rerr 0.046 versus 187.55 and 0.053 for VD3D.
- Single-image 3D scene novel view synthesis on RealEstate10K reaches PSNR 17.15 and LPIPS 0.292, beating ViewCrafter (16.84, 0.341) and ZeroNVS (13.01, 0.448).
- End-to-end scene generation takes about 5 minutes on one A100, reported as 3.2x faster than Cat3D, 1.2x faster than ViewCrafter and 36x faster than ZeroNVS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the dual-branch camera-conditioning mechanism instead of the strongest baseline | Camera-guided video generation · RealEstate10K | FID: 20.89 -> 16.16 | Helped |
| Use the dual-branch camera-conditioning mechanism instead of the strongest baseline | Camera-guided video generation · DL3DV | FID: 20.55 -> 17.74 | Helped |
| Use the dual-branch camera-conditioning mechanism instead of the strongest baseline | Camera-guided video generation · Tanks and Temples | FID: 22.41 -> 19.46 | Helped |
| Use Wonderland instead of ViewCrafter for feed-forward single-image 3D reconstruction | 3D scene generation · RealEstate10K | LPIPS: 0.341 -> 0.292 | Helped |
| Add the ControlNet camera-conditioning branch to a LoRA-only branch (ablation) | Camera-guided video generation · RealEstate10K | FID: 19.02 -> 18.75 | Helped |
| Add the LoRA camera-conditioning branch to the ControlNet-only branch (ablation) | Camera-guided video generation · RealEstate10K | FID: 18.75 -> 17.22 | Helped |
| Keep the 3DVAE encoder frozen rather than finetuning it (ablation) | LaLRM · RealEstate10K | PSNR: 26.14 -> 27.10 | Helped |
| Keep the 3DVAE encoder frozen rather than finetuning it (ablation) | LaLRM · Tanks-and-Temples | PSNR: 20.75 -> 22.66 | Helped |
| Use frozen video latents instead of RGB-14 inputs (ablation) | LaLRM · RealEstate10K | PSNR: 21.39 -> 27.10 | Helped |
| Use frozen video latents instead of RGB-49 inputs (ablation) | LaLRM · Tanks-and-Temples | PSNR: 20.54 -> 22.66 | Helped |
Limitations
- Scene quality degrades when the view range grows large (noted for ViewCrafter; stated limitations of this method are not explicit).
- Video latents restrict reconstruction to the 3DVAE-compressed representation; joint encoder finetuning hurts performance.
Experimental setup
Datasets: RealEstate10K ACID DL3DV-10K DL3DV-140 Tanks-and-Temples Flux.1-generated videos
Benchmarks: RealEstate10K DL3DV-140 Tanks-and-Temples
Metrics: FID FVD Rerr Terr LPIPS PSNR SSIM
Baselines: MotionCtrl VD3D ViewCrafter ZeroNVS WonderJourney Cat3D
Builds on: CogVideoX ControlNet LoRA 3D Gaussian Splatting ViewCrafter
Closest papers by meaning
- Scene Splatter: Momentum 3D Scene Generation from Single Image with Video Diffusion Model (CVPR 2025 · similarity 0.91)
- Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs (CVPR 2025 · similarity 0.91)
- 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors (NeurIPS 2024 · similarity 0.90)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.90)
- BloomScene: Lightweight Structured 3D Gaussian Splatting for Crossmodal Scene Generation (AAAI 2025 · similarity 0.90)
- DiffSplat: Repurposing Image Diffusion Models for Scalable Gaussian Splat Generation (ICLR 2025 · similarity 0.90)
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.