TaleFrame: An Interactive Story Generation System with Fine-Grained Control and Large Language Models

TL;DR

The paper tackles the difficulty of translating user intent into coherent, controllable stories when using large language models. It introduces TaleFrame, an interactive system that decomposes stories into four foundational units (entities, events, relationships, outline) and converts structured inputs into stories via a fine-tuned local Llama-3-8B model trained on a 9,851-entry JSON preference dataset derived from Tinystories, using a JSON2Story pipeline. It presents a detailed interface with drag-and-drop interactions, plus an evaluation including an ablation study and usability study across seven quality dimensions, showing the full unit model achieves superior content control and coherence. The work offers a practical, privacy-conscious approach to co-creative storytelling and provides a dataset and methods that can be adopted by researchers and practitioners for more reliable, user-guided generation.

Abstract

With the advancement of natural language generation (NLG) technologies, creative story generation systems have gained increasing attention. However, current systems often fail to accurately translate user intent into satisfactory story outputs due to a lack of fine-grained control and unclear input specifications, limiting their applicability. To address this, we propose TaleFrame, a system that combines large language models (LLMs) with human-computer interaction (HCI) to generate stories through structured information, enabling precise control over the generation process. The innovation of TaleFrame lies in decomposing the story structure into four basic units: entities, events, relationships, and story outline. We leverage the Tinystories dataset, parsing and constructing a preference dataset consisting of 9,851 JSON-formatted entries, which is then used to fine-tune a local Llama model. By employing this JSON2Story approach, structured data is transformed into coherent stories. TaleFrame also offers an intuitive interface that supports users in creating and editing entities and events and generates stories through the structured framework. Users can control these units through simple interactions (e.g., drag-and-drop, attach, and connect), thus influencing the details and progression of the story. The generated stories can be evaluated across seven dimensions (e.g., creativity, structural integrity), with the system providing suggestions for refinement based on these evaluations. Users can iteratively adjust the story until a satisfactory result is achieved. Finally, we conduct quantitative evaluation and user studies that demonstrate the usefulness of TaleFrame. Dataset available at https://huggingface.co/datasets/guodaosun/tale-frame.

Figures (8)

• Figure 1: The overview of TaleFrame. (a) The Training pipeline comprises three stages: input story text, story structural parsing, and fine-tuning Llama-3-8B. (b) We decompose the story structure parsing with prompt chaining into four steps: extract entities, extract events, extract story outline, and extract relationships. We summarize the four types of human-computer interaction in TaleFrame (c) and map them to the four foundational units (d).
• Figure 2: Interactive patterns. This diagram includes a legend at the bottom right, which explains the symbols used throughout the diagram: Event, Entity, Relationship. ① An internal relationship within a single entity (e.g., self-encouragement). ② A unidirectional relationship between two entities (e.g., entity 1 encourages entity 2). ③ A bidirectional relationship between two entities (e.g., entity 1 and entity 2 encourage each other). ④, ⑤ and ⑥ Represent unidirectional relationships of "single entity to group of multiple entities", "group of multiple entities to single entity" and "group of multiple entities to multiple entities" respectively. ⑦ Complex relationships between entities, either unidirectional or bidirectional, each relationship independently controlled. ⑧ In TaleFrame, the story evolves in linear time and does not support causal reasoning or non-linear narratives.
• Figure 3: Comparison of Prompting Techniques: Box and violin plots show tree edit distances (lower is better) for Zero-shot prompting, TIDD-EC framework, TIDD-EC with CoT, and TIDD-EC with prompt chaining.
• Figure 4: The statistical analysis of the preference dataset.
• Figure 5: An example of story creation using TaleFrame. (a) This diagram has three entities, four events, with clear relationships and a linear evolution of events, which is ideal as an introduction to how to use the TaleFrame. (b) An on-site photo of using TaleFrame.