Research map / 3D vision / 3D generation
Beyond Pixel Histories: World Models with Persistent 3D State
Samuel Garcin, Thomas Walker, Steven McDonagh, Tim Pearce, Hakan Bilen, Tianyu He, Kaixin Wang, Jiang Bian
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 |
|---|---|---|---|
| add persistent 3D state conditioning (vs pixel-history Oasis) (ablation) | Oasis · FVD 600 frames | FVD: 875 -> 148 | Improved |
| add 3D-VAE upscaling of world-frame latents (ablation) | PERSIST · FVD 600 frames | FVD: 247 -> 148 | Improved |
| use larger 3D-XL denoiser (8x spatial tokens) instead of 3D-S (ablation) | PERSIST · FVD 600 frames | FVD: 179 -> 148 | Improved |
| use larger 3D-XL denoiser instead of 3D-S (ablation) | PERSIST · Human study overall | score: 2.6 vs 2.6 | No clear effect |
| use learned camera model instead of ground-truth camera (ablation) | PERSIST · FVD 600 frames | FVD: 152 vs 148 | No clear effect |
| provide ground-truth initial world-frame w0 | PERSIST · FVD 600 frames | FVD: 148 -> 104 | Improved |
| provide ground-truth initial world-frame w0 | PERSIST · Human study overall | score: 2.6 -> 3.0 | Improved |
| use 2 and 4 denoising steps instead of 20 | PERSIST-S · FVD 200/400/600 frames | FVD: 159/170/179 to 207/230/244 | Hurt |
| use 2 and 4 denoising steps instead of 20 | PERSIST-S · inference speed | speedup: 3x improvement; 886 ms (1.13 FPS) | Improved |
| use PERSIST-S instead of WorldMem pixel-memory retrieval | PERSIST-S · time per frame | ms: 6387.1 -> 2672.4 | Improved |
| use PERSIST-S instead of Oasis | PERSIST-S · time per frame | ms: 3007.4 -> 2672.4 | Improved |
| add 3D state to Oasis in human study (PERSIST vs Oasis) | PERSIST · Human study 3D consistency | score: 1.9 -> 2.5 | Improved |
Closest papers by meaning
- Video World Models with Long-term Spatial Memory (NeurIPS 2025 · similarity 0.93)
- WorldMem: Long-term Consistent World Simulation with Memory (NeurIPS 2025 · similarity 0.93)
- WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling (ICML 2026 · similarity 0.92)
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.92)
- Persistent Nature: A Generative Model of Unbounded 3D Worlds (CVPR 2023 · similarity 0.90)
- SimVS: Simulating World Inconsistencies for Robust View Synthesis (CVPR 2025 · 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.
