Research map / 3D vision / 3D generation
WonderWorld: Interactive 3D Scene Generation 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.
Presents WonderWorld, an interactive system that generates connected 3D scenes from a single image in under 10 seconds per scene using layered Gaussian surfels (FLAGS) and training-free guided depth diffusion.
Problem
Existing 3D scene generation is offline, taking tens of minutes to hours per scene because it generates many dense views and optimizes 3D representations, and generated scenes show seams at boundaries, preventing interactive, connected scene generation.
Approach
Generate geometric layers from a single view and inpaint occluded content at the layer level, represent scenes as layered Gaussian surfels initialized from estimated depth and normals (geometry-based initialization) so optimization is only fine-tuning, and steer depth diffusion with partially visible existing depth (guided depth diffusion) to reduce seams between scenes.
Key results
- Generates a scene in 9.5 seconds on an A6000 GPU versus 749.5 (WonderJourney), 798.1 (LucidDreamer), and 766.9 (Text2Room) seconds.
- Human 2AFC bird-eye view preference for WonderWorld is 98.0% to 98.6% against each baseline.
- CLIP score of 29.47 versus 27.34 for the best baseline (WonderJourney); ablation without guidance drops CLIP score to 26.89.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WonderWorld instead of WonderJourney | 3D scene generation · Novel view renderings | CS: 27.34 -> 29.47 | Helped |
| Use WonderWorld instead of LucidDreamer | 3D scene generation · Novel view renderings | CS: 26.72 -> 29.47 | Helped |
| Use WonderWorld instead of Text2Room | 3D scene generation · Novel view renderings | CS: 24.50 -> 29.47 | Helped |
| Use WonderWorld instead of WonderJourney | 3D scene generation · Scene generation on A6000 GPU | time: 749.5 seconds -> 9.5 seconds | Helped |
| Add geometry-based initialization and surfel design (ablation) | WonderWorld · Novel view renderings | CS: 27.23 -> 29.47 | Helped |
| Use three layers instead of a single layer (ablation) | WonderWorld · Novel view renderings | CS: 27.32 -> 29.47 | Helped |
| Add guided depth diffusion (ablation) | WonderWorld · Novel view renderings | CS: 26.89 -> 29.47 | Helped |
Surprising findings
- Single-scene 3D methods like LucidDreamer fail to extrapolate beyond predefined scenes and show severe boundary distortion, and Text2Room's indoor depth inpainting does not generalize to outdoor scenes.
Limitations
- Generated scenes have only frontal-facing surfaces, limiting view synthesis range around the camera.
- Detailed objects such as trees leave holes or floaters when the viewpoint changes.
- Positioned as an interactive prototyping method rather than a full end-to-end solution.
Experimental setup
Datasets: publicly available real images synthetic images generated for testing examples from WonderJourney and LucidDreamer
Benchmarks: custom 28-scene qualitative comparison set (4 test examples x 7 scenes)
Metrics: generation time (seconds) CLIP score CLIP consistency CLIP-IQA+ Q-Align CLIP aesthetic score human 2AFC preference
Baselines: WonderJourney LucidDreamer Text2Room
Builds on: 3D Gaussian Splatting WonderJourney LucidDreamer Marigold Stable Diffusion Inpainting
Closest papers by meaning
- WonderJourney: Going from Anywhere to Everywhere (CVPR 2024 · similarity 0.90)
- Interactive3D: Create What You Want by Interactive 3D Generation (CVPR 2024 · similarity 0.88)
- A 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale (NeurIPS 2026 · similarity 0.88)
- SceneCraft: Layout-Guided 3D Scene Generation (NeurIPS 2024 · similarity 0.87)
- ArtiScene: Language-Driven Artistic 3D Scene Generation Through Image Intermediary (CVPR 2025 · similarity 0.87)
- Wonder3D: Single Image to 3D using Cross-Domain Diffusion (CVPR 2024 · 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.