Look-Before-Move: camera planning with a 50-story 3D benchmark

Look-Before-Move, a camera planning framework published on arXiv on 25 Jun 2026 by Jiaming Bian and seven coauthors, separates what to observe from how to move in dynamic 3D story worlds. The paper defines "Narrative-Grounded World Visual Attention" and implements a three-stage pipeline that converts directorial intent into executable camera motion while respecting geometric and temporal constraints.

What is Look-Before-Move?

Look-Before-Move is a camera planning framework that treats the camera as an embodied observer which must decide what to observe, how to compose the observation, and how to shift attention over time under narrative intent and 3D constraints. The authors frame this capability as Narrative-Grounded World Visual Attention and position it as an alternative to passively interpreting observations, arguing the camera should plan observation before motion.

The paper formalizes the problem for dynamic 3D story worlds and contrasts observation specification with motion execution. The framework components are named explicitly: Semantic Observation Contract, Monte Carlo Viewpoint Search, and Semantic Trajectory Grounding. The submission to arXiv is listed as arXiv:2606.26964 and the paper spans 25 pages with 17 figures.

How does the system work?

Look-Before-Move first converts directorial intent into executable visual constraints with a Semantic Observation Contract, then finds candidate viewpoints with Monte Carlo Viewpoint Search, and finally connects viewpoints into continuous, collision-aware camera motion via Semantic Trajectory Grounding. That three-step breakdown separates the what from the how so the system can prioritize narrative-compliant observations before committing to trajectories.

Semantic Observation Contract encodes the visual requirements implied by narrative intent into semantic constraints the renderer or planner can follow. Monte Carlo Viewpoint Search samples and evaluates viewpoints against those constraints to find narrative-compliant and geometrically feasible camera positions. Semantic Trajectory Grounding links chosen viewpoints into smooth, temporally coherent, and collision-aware camera trajectories suitable for animated scenes.

How was it evaluated?

The authors built a dynamic 3D Story World Benchmark based on StoryBlender covering 50 stories, 457 scenes, and 1,585 shots to test the framework in environments with animated characters and semantic scene configurations. Experiments reported in the paper show improvements in subject perception, intent consistency, and trajectory quality compared with representative baselines.

The benchmark is a concrete artifact of the work: 50 stories produce a total of 457 scenes and 1,585 shots, all embedded in executable 3D environments drawn from StoryBlender. Those dataset counts are the primary quantitative signals given in the paper; the submission metadata also notes the document length and figure count mentioned above.

Why it matters

Look-Before-Move pushes camera planning toward explicit attention planning, a shift that matters for embodied AI agents, virtual cinematography, and interactive storytelling. By formalizing observation specification separately from motion execution, the approach can enforce narrative goals while keeping trajectories physically plausible, which is relevant for systems that must satisfy both directorial intent and real-world constraints. The StoryBlender-based benchmark supplies a standardized testbed to compare future methods on narrative compliance and trajectory quality.

What to watch

Look-Before-Move's next milestones are external adoption of the StoryBlender benchmark and independent replication of the reported gains in subject perception, intent consistency, and trajectory quality. Watch for follow-up code, data releases, or workshop papers that apply the Semantic Observation Contract and Monte Carlo Viewpoint Search in other interactive or embodied settings.

References and provenance: the framework and dataset counts appear in the arXiv submission "Look-Before-Move: Narrative-Grounded World Visual Attention in Dynamic 3D Story Worlds", arXiv:2606.26964, submitted 25 Jun 2026 by Jiaming Bian and seven coauthors. The paper length is listed as 25 pages with 17 figures.