Research map / 3D vision / 3D generation
Look-Before-Move: Narrative-Grounded World Visual Attention in Dynamic 3D Story Worlds
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 Look-Before-Move, a multi-agent camera planning framework that first builds a semantic observation contract, then searches viewpoints via Monte Carlo sampling and grounds them into smooth trajectories, evaluated on a new executable 3D story-world benchmark built on StoryBlender.
Problem
Camera planning in dynamic 3D story worlds requires deciding what visual evidence to acquire from narrative intent while respecting occlusion, collision, and smooth motion; existing methods assume the observation objective is already specified.
Approach
Separates observation specification (semantic observation contract from directorial intent) from motion execution: Monte Carlo viewpoint search with tournament reranking and render-based reflection selects narrative-compliant, feasible viewpoints before a trajectory is grounded through them with collision-aware smoothing and temporal editing.
Key results
- Look-Before-Move achieves Overall 71.70 versus 65.62 for the strongest baseline CCD (+6.08) and 59.64 for FilmAgent-adapted.
- Removing Monte Carlo search (w/o MLS) drops Overall from 71.70 to 60.52.
- Human IC rating of 4.09 for Look-Before-Move vs 3.25 for CCD, 3.10 for GenDoP, and 2.53 for Director, from 95 participants.
Surprising findings
- Raising the candidate count beyond 16 yields only marginal gains (16 to 50 candidates adds 0.402 Overall), suggesting gains are not explained by pose-pool size alone.
- Qwen2.5-VL-7B IC3 scores favor CCD over Look-Before-Move (24.86 vs 23.26), while GPT-4o, Gemini, and Claude rank Look-Before-Move first.
Limitations
- Evaluation relies on automatic metrics built on YOLO11x, ORB, and VLM judges, with a user study of limited size.
- Runtime is higher than CCD (26.50 vs 18.16 seconds per effective video second).
- Evaluation focuses on short story segments; long-horizon constraints such as screen direction and story-level memory are not addressed.
Experimental setup
Datasets: StoryBlender-based 3D Story World Benchmark
Benchmarks: 3D Story World Benchmark
Metrics: subject perception (SP1-3) intent consistency (IC1-3) trajectory quality (TQ1-3) user study score overall
Baselines: CCD GenDoP+RGBD GenDoP+noRGBD FilmAgent-adapted ET+Director A/B/C
Builds on: StoryBlender FilmAgent GenDoP Reflexion
Closest papers by meaning
- LogiStory: A Logic-Aware Framework for Multi-Image Story Visualization (ICLR 2026 · similarity 0.87)
- DriveDreamer-Policy: A Geometry-Grounded World-Action Model for Unified Generation and Planning (NeurIPS 2026 · similarity 0.86)
- DynamicVerse: A Physically-Aware Multimodal Framework for 4D World Modeling (NeurIPS 2025 · similarity 0.86)
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.86)
- MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (NeurIPS 2026 · similarity 0.86)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.86)
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.