Research map / 3D vision / 3D generation
Persistent Nature: A Generative Model of Unbounded 3D Worlds
Lucy Chai, Richard Tucker, Zhengqi Li, Phillip Isola, Noah Snavely
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a persistent global scene representation rather than autoregressive frame prediction | Persistent Nature · InfiniteNature-Zero trajectory, 100 steps forward | FID: 28.15 -> 26.09 | Improved |
| Use a persistent global scene representation rather than autoregressive frame prediction | Persistent Nature · InfiniteNature-Zero trajectory, forward-backward cycle | cycle consistency L1 error: 3.94 -> 0.00 | Improved |
| Use a persistent global scene representation rather than autoregressive frame prediction | Persistent Nature · InfiniteNature-Zero trajectory, one step | consistency L1 error: 1.84 -> 2.12 | Hurt |
| Increase layout resolution and ray samples (ablation) | Layout generator · 32×32 rendered images | FID: 33.66 -> 16.06 | Improved |
| Add depth and sky-mask outputs to refinement (ablation) | Refinement network · Training camera poses | FID: 26.30 -> 23.75 | Improved |
| Add depth and sky-mask outputs to refinement (ablation) | Refinement network · Random camera poses | FID: 27.08 -> 27.25 | No clear effect |
| Add depth and sky-mask outputs to refinement (ablation) | Refinement network · All camera poses | consistency error: 5.08 -> 5.81 | Hurt |
| Add projected 3D world-space noise to refinement (ablation) | Layout refinement network · Training camera poses | FID: 23.75 -> 21.42 | Improved |
| Add projected 3D world-space noise to refinement (ablation) | Layout refinement network · Random camera poses | FID: 27.25 -> 23.39 | Improved |
| Add projected 3D world-space noise to refinement (ablation) | Layout refinement network · All camera poses | consistency error: 5.81 -> 3.91 | Improved |
| Replace the extended layout representation with extended triplane planes (ablation) | Persistent Nature · Training camera poses | FID: 21.42 -> 24.47 | Hurt |
| Replace the extended layout representation with extended triplane planes (ablation) | Persistent Nature · Forward camera motion | FID: 26.67 -> 34.89 | Hurt |
Closest papers by meaning
- Beyond Pixel Histories: World Models with Persistent 3D State (ICML 2026 · similarity 0.90)
- WorldFlow3D: Flowing Through 3D Distributions for Unbounded World Generation (ECCV 2026 · similarity 0.90)
- WorldGrow: Generating Infinite 3D World (AAAI 2026 · similarity 0.90)
- 3D-SceneDreamer: Text-Driven 3D-Consistent Scene Generation (CVPR 2024 · similarity 0.88)
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.87)
- SimVS: Simulating World Inconsistencies for Robust View Synthesis (CVPR 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.
