StateFlow: State-Centric Generative Previsualization
Paper: arXiv:2508.03421 Authors: Yuyang Yin, Zixiang Li, Longxuan Deng Date: 2026-08-13
Summary
Previsualization (previs) is an intermediate step between creative ideation and final production in film, games, architecture, and urban design. It allows creators to iteratively refine scenes, actions, camera shots, and spatiotemporal dynamics. Existing generative approaches use simple text prompts and attempt to jointly control all these factors through one-shot image or video synthesis, which results in weak controllability and limited support for iterative editing.
The authors observe that a world fundamentally consists of multiple elements (geometry, appearance, and other attributes) plus a camera. Different frames are produced by locally modifying or recombining portions of this shared state, while the rest is largely reused. The missing component is therefore an explicit, persistent working state.
StateFlow Framework
StateFlow is a state-centric generative previsualization framework. Instead of generating a video in one pass, it uses an editable 3D world to organize scene structure, evolution, and the camera. Off-the-shelf video models are applied on top to boost visual fidelity when higher realism is required.
The 3D world is maintained as a persistent, structured 3D state of scene elements and camera configurations, serving as the core working representation for previsualization. StateFlow builds, evolves, and accesses this world state in three stages:
1. State Construction — Generated 2D content is lifted into a coherent 3D world via *prior-guided, conflict-aware dual-view initialization*. 2. State Evolution — User intent is translated into structured state transitions while preserving world memory, avoiding full scene regeneration for every edit. 3. State Access — Camera plans are refined into visually feasible trajectories using *rendering feedback reflection*, instead of relying solely on VLM-based semantics.
Results
Experiments demonstrate that StateFlow can generate high-quality 3D worlds for video creation and game-like prototyping, supporting iterative scene editing and camera refinement that is difficult to achieve with one-shot video synthesis.
Tags
computer-vision, generative-ai, 3d-reconstruction, previsualization, scene-editing, video-generation, camera-planning