Research map / 3D vision / 3D generation
Fast SceneScript: Fast and Accurate Language-Based 3D Scene Understanding via Multi-Token Prediction
Ruihong Yin, Xuepeng Shi, Oleksandr Bailo, Marco Manfredi, Theo Gevers
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 multi-token prediction heads (n=10) to SceneScript without filtering | SceneScript · ASE layout estimation (test) | mean F1-Score: 0.915 -> 0.814 | Hurt |
| add self-speculative decoding with parameter sharing to MTP (n=10) | SceneScript + MTP · ASE layout estimation (test) | mean F1-Score: 0.814 -> 0.912 | Improved |
| use Fast SceneScript (SSD, n=10) instead of next-token prediction | SceneScript · ASE layout estimation (val) | latency: 382 ms -> 75 ms | Improved |
| add MTP with SSD to SceneScript | SceneScript · Structured3D layout estimation (test) | mean F1-Score: 0.774 -> 0.791 | Improved |
| loosen numerical token check to tau=2 in SSD (ablation) | Fast SceneScript (SSD) · ASE layout estimation (val), n=8 | accepted tokens per inference: 6.53 -> 7.46 | Improved |
| increase tau from 2 to 5 (ablation) | Fast SceneScript (SSD) · ASE layout estimation (val), n=8 | mean F1-Score: 0.912 -> 0.911 | No clear effect |
| use CGD confidence head instead of ProductThre scoring (ablation) | Fast SceneScript · ASE layout estimation (val), n=8 | accepted tokens per inference: 4.67 -> 6.29 | Improved |
| use CGD instead of SoftmaxThre scoring (ablation) | Fast SceneScript · ASE layout estimation (val), n=8 | latency: 116 ms -> 91 ms | Improved |
| add shared token heads with projection block (head sharing) to SSD (ablation) | Fast SceneScript (SSD) · ASE layout estimation (val), n=8 | decoder parameters: 23.67 M -> 15.05 M | Improved |
| add head sharing to CGD (ablation) | Fast SceneScript (CGD) · ASE layout estimation (val), n=8 | decoder parameters: 31.03 M -> 16.10 M | Improved |
| add head sharing to CGD (ablation) | Fast SceneScript (CGD) · ASE layout estimation (val), n=8 | mean F1-Score: 0.910 -> 0.911 | No clear effect |
| use Fast SceneScript (SSD) instead of SceneScript + MTP | SceneScript · SceneCAD layout estimation (val) | F1-Score: 0.460 -> 0.556 | Improved |
Closest papers by meaning
- SceneScaffold: Active Scene-State Construction for Unified 3D Scene Understanding (NeurIPS 2026 · similarity 0.88)
- The Scene Language: Representing Scenes with Programs, Words, and Embeddings (CVPR 2025 · similarity 0.88)
- Lexicon3D: Probing Visual Foundation Models for Complex 3D Scene Understanding (NeurIPS 2024 · similarity 0.88)
- Know Your Neighbors: Improving Single-View Reconstruction via Spatial Vision-Language Reasoning (CVPR 2024 · similarity 0.87)
- SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting (NeurIPS 2025 · similarity 0.87)
- SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes (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.
