DiaryPlay: Interactive Narrative Authoring
- DiaryPlay is an AI-assisted authoring system that transforms a single-branch natural language story into an interactive branch-and-bottleneck vignette.
- It uses a three-part specification—environment, characters, and events—to automatically generate and refine interactive narrative content.
- The system employs a Controlled Divergence module that balances authorial intent with viewer interaction to maintain narrative coherence.
Searching arXiv for DiaryPlay and closely related diary/interactive-vignette work to ground the article in current papers. DiaryPlay is an AI-assisted authoring system for interactive vignette creation in everyday storytelling. It takes a natural language story as input, extracts the three core elements of an interactive vignette—environment, characters, and events—and automatically transforms a single-branch story input into a branch-and-bottleneck structure using an LLM-powered narrative planner. In this formulation, a viewer role-plays a player character (PC), interacts with non-player characters (NPCs), and influences the narrative in an interactive environment, while the authored storyline remains the governing structure (Xu et al., 15 Jul 2025).
1. Definition, scope, and design rationale
DiaryPlay addresses a specific mismatch between interactive narrative systems and everyday storytelling. Interactive vignettes are described as a popular and immersive visual storytelling approach, but they have not been widely used by everyday storytellers because authoring requires scene design, object placement, character scripting, and branching logic. That complexity conflicts with the immediacy of everyday storytelling, which is ordinarily lightweight and usually expressed as text, photos, or video rather than as a branching system (Xu et al., 15 Jul 2025).
The system’s explicit objective is therefore twofold: to transform a single-branch story into something interactive, and to keep authors in control by letting them edit and refine system-generated output. A formative interview with 13 participants motivated these design goals. The resulting workflow is mixed-initiative rather than fully automatic: generative AI bootstraps the interactive vignette, and authors refine the generated environment, characters, and events instead of constructing them from scratch (Xu et al., 15 Jul 2025).
This distinguishes DiaryPlay from adjacent diary-oriented systems. Automatic diary generation in human-robot interaction has been used to verbalize joint experiences as diary text (Ichikura et al., 2023), elicitation diary studies have used automatic contextual cue generation to support later recall (Li et al., 2024), and robots have been evaluated as interactive diary instruments for in-home data collection (Xu et al., 8 Jan 2025). DiaryPlay instead treats the diary-like story as source material for an interactive vignette rather than as a memory cue, a self-report instrument, or a robot-authored recollection (Xu et al., 15 Jul 2025).
2. Interactive vignette representation
DiaryPlay adopts a three-part specification for interactive vignettes: environment, characters, and events. This decomposition is central because it exposes the structural gap between an author’s short natural-language story and the richer representation needed for an interactive system (Xu et al., 15 Jul 2025).
| Element | Representation | Key fields or form |
|---|---|---|
| Environment | Scene layout and objects | location, scale, available actions, interaction zone type |
| Characters | PC and NPC descriptions | name, age, personality, social role, mood, language style, conversation examples, sprites/visuals |
| Events | Ordered sequence of key events | one or more simultaneous activities as |
The environment comprises the scene layout and its objects. Each object is represented with fields such as location, scale, available actions, and interaction zone type. The character layer separates the first-person narrator, interpreted as the PC, from other named or implied people, who become NPCs. The event layer models the story as an ordered sequence of key events, where each event contains one or more simultaneous activities represented as (Xu et al., 15 Jul 2025).
This representation is neither a free-form transcript nor a conventional branching script. It is closer to a structured operationalization of story content. A plausible implication is that DiaryPlay treats everyday narrative as a schedulable interactive system: spatialized by the environment, socialized by the character model, and temporally ordered by key events.
3. Authoring pipeline from story to vignette
The end-to-end pipeline begins with a natural-language story and ends with a playable interactive vignette. The authoring component starts with an LLM-based extractor that derives environment, characters, and events from the story text (Xu et al., 15 Jul 2025).
Environment construction is split into two steps. First, the LLM infers a high-level room layout using layout templates based on urban-space categories and labels the rooms. The author can revise labels such as bedroom, kitchen, or living room. Second, once the layout is confirmed, the system fills rooms with objects using a greedy algorithm. These objects include event-related objects, environment-related objects, and decorative objects if space remains. Placement uses LLM reasoning about object-object relations, object-room relations, and paths or accessibility. For each object, the system also infers possible actions and an interaction trigger zone type. The named trigger zone categories are on, partial, around, and directional; examples include sleeping on a bed as on, sitting partially on a sofa as partial, sitting around a table as around, and opening a fridge from the front as directional (Xu et al., 15 Jul 2025).
Character extraction identifies the first-person pronoun, such as “I,” as the PC and maps other named or implied people to NPCs. The system extracts explicitly stated attributes such as name or social role, but intentionally leaves some attributes blank if the story does not specify them. The character panel also includes conversation simulation, allowing the author to talk to an NPC and inspect LLM-generated responses in that character’s voice. The paper emphasizes that the system avoids over-inference and does not aggressively hallucinate character traits not specified in the story (Xu et al., 15 Jul 2025).
Event extraction proceeds in three steps. The LLM first extracts activities for each character in the form . The system then groups simultaneous activities into a single key event and finally orders the key events chronologically. This preserves the temporal flow of the story while yielding a runtime structure that can support interaction (Xu et al., 15 Jul 2025).
The authoring interface mirrors this decomposition through three panels: Environment, Characters, and Events. The illustrated workflow includes changing room labels, moving objects, adding missing objects, editing character personas, using simulated conversation to refine NPC voice, and correcting event grouping or object assignment. The system is therefore not described as a one-shot generator of finished artifacts, but as a low-effort, human-in-the-loop workflow (Xu et al., 15 Jul 2025).
4. Controlled Divergence and branch-and-bottleneck narrative control
The system’s main technical contribution is the Controlled Divergence (CD) module, an LLM-powered narrative planner that turns a single-branch story into a branch-and-bottleneck narrative. In this structure, key events are fixed checkpoints, the viewer may take divergent actions between them, and the story then converges back to the next key event. The paper characterizes this as a good fit for everyday storytelling because authors typically think in terms of a single linear story rather than explicit multi-branch planning (Xu et al., 15 Jul 2025).
Controlled Divergence operates as a two-stage loop. In the first stage, to maintain real-time responsiveness, the system plans NPC behavior ahead of time. For each NPC it generates two next-activity plans: one if the PC follows the next key event, and one if the PC diverges. This planning is conditioned on the NPC persona, the storyline, past activities, current ongoing activities, and the next key event. In the second stage, when the PC acts, the module checks whether the player followed the expected event and selects the appropriate preplanned NPC action. After the NPC acts, the system returns to the planning stage (Xu et al., 15 Jul 2025).
The system also includes two guidance mechanisms. If the PC diverges from the next key event, the LLM generates an inner voice thought bubble that nudges the viewer, for example, “Time to enjoy a meal!” If the viewer deviates too far from the authored storyline, NPC chat steers them back; the example given is an NPC replying, “Dinner is important. Let’s have dinner together.” These mechanisms are designed to preserve authorial intent without making the guidance forceful (Xu et al., 15 Jul 2025).
A common misconception is that DiaryPlay simply generates open-ended branches. The paper instead presents bounded divergence: freedom exists between key events, but the authored event order remains the controlling scaffold. Another misconception is that the system replaces authorial design. The paper explicitly frames AI as a collaborator rather than an autopilot and argues that the relevant design problem is balancing automation with preserved author agency (Xu et al., 15 Jul 2025).
5. Implementation and empirical results
The implementation uses Phaser, map templates created with Tiled, visual assets from open resources and DALL·E 2, and the LLM gpt-4o-2024-05-13. The current version supports a maximum of 3 characters. The system also monitors dialogue for policy violations and withholds harmful responses (Xu et al., 15 Jul 2025).
The technical evaluation examined the believability of NPC activities generated by Controlled Divergence. Sixteen participants authored interactive vignettes, another user interacted with each vignette to trigger the CD module, and a separate group of 16 evaluators ranked believability. Each evaluator ranked 12 sets of activity tables, yielding 192 sets of rankings. Five conditions were compared: CD, BL, HA, PO, and SO. Believability was ranked from 1 as best to 5 as worst. The average rankings were CD: 2.37, HA: 2.58, PO: 2.74, SO: 3.31, and BL: 4.00. Friedman’s test reported . Post-hoc Nemenyi tests showed CD significantly better than BL, CD not significantly different from HA, PO and SO both better than BL, CD better than SO, and CD versus PO not significantly different. The paper interprets these results as showing that DiaryPlay-generated character activities are better than baseline, comparable to human-authored behavior, and supported by both persona consistency and narrative coherence; persona-only was better than storyline-only in average ranking (Xu et al., 15 Jul 2025).
The user study involved 16 participants formed into 8 pairs, with each session lasting about 1.5 hours and compensation of 30,000 KRW. Participants authored a vignette, viewed each other’s vignettes, reported takeaway messages, discussed whether viewer interpretation matched author intent, and were interviewed; one solo participant viewed a vignette from the technical evaluation instead. The analysis used thematic analysis. Participants completed their vignettes in 12–19 minutes, with an average of 16 minutes, and generally regarded this as appropriate for everyday storytelling. The system correctly extracted all characters in all 16 stories. Participants reported that the environment panel’s automatic object generation and placement provided a strong starting point, the characters panel was especially useful because many found personality construction difficult, and the events panel generally structured events correctly, with mismatches arising mainly from vagueness in input stories and being easy to fix by adding an activity or editing an event directly (Xu et al., 15 Jul 2025).
For viewing, all authors felt their intended storyline was conveyed accurately. Even when viewers diverged from the main path, they still understood the main takeaways. The paper attributes this to three supports: the text caption as a stable reference, inner voice guidance, and glowing objects signaling the next key event. Participants described the experience as immersive, lively, and fun to replay, and many replayed the vignette multiple times to explore different interactions (Xu et al., 15 Jul 2025).
6. Position within diary and storytelling research
DiaryPlay occupies a boundary zone between diary-derived narrative input and interactive storytelling output. Its input is a short everyday story rather than a manually authored interactive script, which aligns it with diary-like modes of capture, but its output is an interactive vignette rather than a diary entry, a recall aid, or a study instrument (Xu et al., 15 Jul 2025).
This position becomes clearer when contrasted with adjacent systems. In human-robot interaction, an automatic diary generation system has used interaction and dialogue history to produce a diary from past joint experiences; in evaluation, diaries with information on joint experiences were selected higher than diaries without such information because they showed more cooperation and more intimacy from the robot (Ichikura et al., 2023). DiaryHelper, by contrast, is an automatic contextual information recording agent for elicitation diary studies based on episodic memory theory; it records five contextual dimensions—time, location, emotion, people, and activity—and improved total recall score from 5.20 to 6.63 with in a within-subject study over two weeks (Li et al., 2024). A separate line of work has explored the Diary Robot system, in which a robot plays the role of an interactive diary; that system used six predetermined questions, could ask LLM-generated follow-up questions, and achieved 87.5% on-time submissions without reminders in the robot condition (Xu et al., 8 Jan 2025).
These systems share a concern with diary mediation, but they solve different problems. The robot diary work targets data collection, elicitation, and social interaction in naturalistic settings (Xu et al., 8 Jan 2025); the automatic diary generation work targets verbalization of joint experiences for favorability and relationship building (Ichikura et al., 2023); the contextual recording work targets cue enrichment for later recollection (Li et al., 2024). DiaryPlay targets authoring complexity in interactive storytelling and uses a natural-language story as a specification seed rather than as a log to preserve or analyze (Xu et al., 15 Jul 2025).
A broader theoretical parallel appears in research on online diary self-representation. Work on Livejournal and Touchgraph argues that digital diary practice is shaped by interface, functionality, visibility, and intersubjectivity, and that the diarist moves from the lonely space of writing to the community space of publication through a strategy of automédiation (Georges, 2017). This suggests a useful conceptual lens for DiaryPlay: the system converts private or semi-private everyday narration into a shareable, role-playable, socially legible vignette. That implication is interpretive rather than explicitly stated in the DiaryPlay paper, but it is consistent with the broader literature on mediated self-representation and interactive narrative performance (Georges, 2017).
7. Limitations, generalization, and prospective directions
The paper notes several limitations. Environment generation can still be improved, object orientation is not modeled, and future 3D environments would require better spatial prompting. The evaluation scope is limited because believability is only one metric; future work is proposed to measure diversity, creativity, long-term usage, and usability quantitatively. The current implementation supports only three characters. The paper also suggests that DiaryPlay may be useful beyond everyday storytelling, including mental health counseling and educational storytelling, while noting that those contexts may require stricter control over divergence (Xu et al., 15 Jul 2025).
These limitations help delimit the present system. DiaryPlay is not presented as a general interactive fiction engine, a free-form multi-character simulator, or a replacement for authorial scripting at arbitrary scale. Its demonstrated contribution is narrower and technically specific: it lowers the authoring barrier for everyday storytellers by extracting environment, characters, and events from a short text story, then using Controlled Divergence to preserve the main storyline while allowing bounded viewer interaction (Xu et al., 15 Jul 2025).
Within the current research landscape, this positions DiaryPlay as a mixed-initiative narrative authoring system that operationalizes everyday storytelling as structured, role-playable interaction. A plausible implication is that its long-term significance will depend less on raw generative breadth than on how well it maintains persona, narrative coherence, and authorial intent under increasingly complex viewer behavior. That emphasis is already visible in the reported result that persona-only generation outperformed storyline-only generation in average believability ranking, suggesting that future systems may benefit from stronger persona construction and maintenance over time (Xu et al., 15 Jul 2025).
