EvoStory: a model-agnostic "blueprint-to-render" framework for long-form narrative consistency
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
In recent years, diffusion models have achieved significant progress in single-image generation. However, they still face a 'semantic gap' between linear text and two-dimensional spatial consistency in long-form narrative image generation, leading characters to easily fall into failure modes of 'identity erasure' and 'semantic overriding'. Existing technologies mostly rely on implicit attention or temporary memory, lacking explicit structural control and cross-cycle asset reuse capabilities, and are highly coupled with underlying models. To address these challenges, this paper proposes EvoStory, a Model-Agnostic structured visual narrative generation framework. We innovatively introduce a Narrative-to-Blueprint protocol, transforming ambiguous natural language into an intermediate representation layer containing spatial layouts and high-fidelity character matrices, acting as a Constraint Injector Middleware. The core of this framework adopts a Dual Decoupling strategy: in the temporal dimension, separating characters' inherent identity from stage-specific features to support long-cycle character dynamic evolution; in the logical dimension, using region masks and geometric skeletons to completely separate narrative structure from visual rendering, eradicating feature collapse and fusion issues in complex multi-character interactions. Experiments show that EvoStory can not only precisely drive local SDXL workflows to achieve absolute spatial hard control, but also serve as a universal Visual Prompting compiler, significantly suppressing prior collapse and spatial hallucinations in closed-source SOTA multimodal large models like Gemini during long-form narratives. This work achieves complete decoupling of Planner and Painter, establishing a solid system architecture foundation for cross-cycle consistent high-fidelity storyboard generation, providing an industrially valuable solution.
Fields of Study & Subjects
Generative Adversarial Networks and Image Synthesis
Multimodal Machine Learning Applications
Digital Media and Visual Art
Publication details
- Journal / source
- DR-NTU (Nanyang Technological University)
- Publication date
- 1 January 2026
- Language
- EN