Persistent Story World Simulation with Continuous Character Customization

Abstract

Story visualization has gained increasing attention in computer vision. However, current methods often fail to achieve a synergy between accurate character customization, semantic alignment, and continuous integration of new identities. To tackle this challenge, in this paper we present EverTale, a story world simulator for continuous story character customization. We first propose an All-in-One-World Character Integrator to achieve continuous character adaptation within unified LoRA module, eliminating the need for per-character optimization modules of previous methods. Then, we incorporate a Character Quality Gate via MLLM-as-Judge to ensure the fidelity of each character adaptation process through chain-of-thought reasoning, determining whether the model can proceed to the next character or require additional training on the current one. We also introduce a Character-Aware Region-Focus Sampling strategy to address the identity degradation and layout conflicts in existing multi-character visual storytelling, ensuring natural multi-character generation by harmonizing local character-specific details with global scene context with higher efficiency. Experimental results show that our EverTale achieves superior performance against a wider range of compared methods on both single- and multi-character story visualization. Codes will be available.

Figures (21)

• Figure 1: EverTale achieves efficient adaptation to new characters effectively preserving the identity of learned ones, e.g., the high-fidelity preservation of V2 when proceeding to V6 on Pororo ((b).$1^{st}$ row) and of V2 when integrating V4 on Frozen ((b).$2^{nd}$ row).
• Figure 2: The overall framework of EverTale. (a).New characters can be continually learned using the All-in-One World Character Integrator. (b).Character Quality Gate ensures per-character customization performance. (c).Character-Aware Region-Focused Sampling strategy achieve accurate multi-character story visualization.
• Figure 3: Visualization comparison for single and multiple character, where the arrows indicate the character learning sequence for CL-based methods. Our method excels in character identity preservation and text-semantic alignment in both story visualization situations.
• Figure 4: The ablation visualizations of the effectiveness of different designs in EverTale.
• Figure 5: The complete evaluation prompt generation instruction $T_e$ used in the Character Quality Gate via MLLM-as-Judge for guiding the multimodal large language model (MLLM) in generating a suitable evaluation prompt for EverTale.