English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

StateFlow: A State-Centric Generative Previsualization Framework with Editable 3D World States (arXiv 2508.03421)

Forum topic · 小凯 · 2026-08-14

Summary

StateFlow (arXiv:2508.03421) is a state-centric generative previsualization framework from researchers Yuyang Yin, Zixiang Li, and Longxuan Deng in the computer vision field. Previsualization bridges ideas and production in film, games, architecture, and urban planning, but existing generative methods rely on simple prompts and one-shot image or video synthesis, offering weak controllability and limited iterative editing. StateFlow's key insight is that different frames derive from local modifications of a shared, persistent world state. The framework maintains an explicit, editable 3D world state—structured scene elements and camera configurations—as its core working representation, optionally enhanced by off-the-shelf video models for higher fidelity. It operates in three stages: state construction lifts generated 2D content into a coherent 3D world via prior-guided, conflict-aware dual-view initialization; state evolution translates user intent into structured state transitions while preserving world memory to avoid full-scene regeneration; and state access refines camera plans into visually feasible trajectories using render-feedback reflection rather than relying solely on VLM semantics. Experiments show StateFlow generates high-quality 3D worlds for video creation and game-like prototyping.

StateFlow: Building, Evolving, and Accessing 3D World States for Previsualization

Paper Overview

  • Field: Computer Vision (CV)
  • Authors: Yuyang Yin, Zixiang Li, Longxuan Deng
  • arXiv: 2508.03421

Abstract (Translated Summary)

Previsualization is the intermediate layer between ideas and production in film, games, architecture, and urban planning. It allows creators to iteratively refine scenes, actions, camera work, and spatiotemporal dynamics. However, existing generative methods rely on simple prompts, jointly controlling all these factors through one-shot image or video synthesis — resulting in weak controllability and limited support for iterative editing.

Fundamentally, a world consists of multiple elements with geometry, appearance, and other attributes, plus a camera. Different frames are produced by local modifications or recombinations of this shared state, with the rest largely reused. The authors argue the missing component is an explicit and persistent working state.

To this end, the paper proposes StateFlow, a state-centric generative previsualization framework. Instead of one-shot video generation, StateFlow organizes scene structure, evolution, and cameras with an editable 3D world, while leveraging off-the-shelf video models to enhance visual effects when higher fidelity is needed.

The world is maintained as a persistent, structured 3D state of scene elements and camera configurations — the core working representation for previsualization. StateFlow builds, evolves, and accesses this world state through three stages:

1. State Construction: lifts generated 2D content into a coherent 3D world via prior-guided, conflict-aware dual-view initialization. 2. State Evolution: translates user intent into structured state transitions while preserving world memory, avoiding full-scene regeneration on every edit. 3. State Access: refines camera plans into visually feasible trajectories using render-feedback reflection, rather than relying solely on VLM semantics.

Experiments demonstrate that StateFlow can generate high-quality 3D worlds for video creation and game-like prototyping.

Original Abstract (Excerpt)

> Previsualization is an intermediate layer between ideas and production in film, games, architecture, and urban design...

Full text: arXiv:2508.03421

---

*Auto-collected on 2026-08-14.*

Tags

#stateflow#previsualization#3d-world-models#generative-ai#computer-vision#video-generation#scene-editing#arxiv

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633451