Research map / 3D vision / 3D generation
WonderJourney: Going from Anywhere to Everywhere
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.
Proposes WonderJourney, a training-free modular framework that uses an LLM, a text-driven point-cloud generator and a VLM validator to produce long journeys of coherently connected 3D scenes from a text or image start.
Problem
Perpetual 3D scene generation: producing a long sequence of diverse yet geometrically and semantically coherent 3D scenes starting from any user-provided text or image, unlike prior perpetual view generation restricted to a single scene type.
Approach
Auto-regressive LLM-generated scene descriptions drive a text-guided outpainting pipeline that lifts each image to a colored point cloud, with depth refinement (SAM segment-based frontal planes, sky handling), depth-alignment and occlusion re-rendering consistency, and VLM-based rejection and regeneration of artifacts.
Key results
- Human study with 400 participants: WonderJourney preferred over InfiniteNature-Zero at 92.7% (diversity), 94.9% (visual quality), 91.5% (complexity), 88.6% (overall).
- Versus SceneScape, preferred at 88.8% (diversity), 83.4% (quality), 80.0% (complexity), 90.3% (overall).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WonderJourney instead of InfiniteNature-Zero | WonderJourney · InfiniteNature-Zero comparison | diversity preference: 92.7% | Helped |
| Use WonderJourney instead of InfiniteNature-Zero | WonderJourney · InfiniteNature-Zero comparison | visual quality preference: 94.9% | Helped |
| Use WonderJourney instead of InfiniteNature-Zero | WonderJourney · InfiniteNature-Zero comparison | scene complexity preference: 91.5% | Helped |
| Use WonderJourney instead of InfiniteNature-Zero | WonderJourney · InfiniteNature-Zero comparison | overall interestingness preference: 88.6% | Helped |
| Use WonderJourney instead of SceneScape | WonderJourney · SceneScape comparison | diversity preference: 88.8% | Helped |
| Use WonderJourney instead of SceneScape | WonderJourney · SceneScape comparison | visual quality preference: 83.4% | Helped |
| Use WonderJourney instead of SceneScape | WonderJourney · SceneScape comparison | scene complexity preference: 80.0% | Helped |
| Use WonderJourney instead of SceneScape | WonderJourney · SceneScape comparison | overall interestingness preference: 90.3% | Helped |
Surprising findings
- Text-guided outpainters tend to avoid generating partial objects, so leaving too little empty space in the rendered image leads to mere extrapolation rather than new objects.
Limitations
- Evaluation relies on a small set of mixed self-collected and generated inputs and a human preference study rather than standard benchmarks.
- Persistent point-cloud representation is replaced by image formulation due to GPU memory cost for long trajectories.
- Depth estimator is not geometry-aware, requiring heuristic refinements such as sky and far-plane cutoffs.
Experimental setup
Datasets: photos taken by authors copyright-free online photos generated examples from DALL-E 3
Metrics: human preference rate on diversity visual quality scene complexity overall interestingness
Baselines: InfiniteNature-Zero SceneScape
Builds on: Infinite Nature SceneScape Stable Diffusion MiDaS SAM
Closest papers by meaning
- WonderWorld: Interactive 3D Scene Generation from a Single Image (CVPR 2025 · similarity 0.90)
- 3D-SceneDreamer: Text-Driven 3D-Consistent Scene Generation (CVPR 2024 · similarity 0.86)
- BloomScene: Lightweight Structured 3D Gaussian Splatting for Crossmodal Scene Generation (AAAI 2025 · similarity 0.85)
- SemCity: Semantic Scene Generation with Triplane Diffusion (CVPR 2024 · similarity 0.85)
- A 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale (NeurIPS 2026 · similarity 0.85)
- CymbaDiff: Structured Spatial Diffusion for Sketch-based 3D Semantic Urban Scene Generation (NeurIPS 2025 · similarity 0.84)
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.