Narrative Keyframing for Generative Creative Writing

The paper introduces narrative keyframing, a new interaction technique for AI-assisted creative writing that lets writers specify different types of narrative constraints at selected moments in a story, then use AI to generate intervening prose. Inspired by keyframing in animation, the approach offers a flexible way to connect story planning with adaptive control over generated text.

The paper explores three types of keyframes:

• Plot keyframes define significant events in a story.

• Character keyframes represent how individual characters change over the narrative.

• Perspective keyframes capture how individual characters experience different events through first-person narratives.

Plot and character keyframes offer a flexible way to adapt high-level conditioning to more customizable, iterative, and fine-scale control, while perspective keyframes add a new way to control characterization and focalization by using first-person narratives as an intermediary.

The system is organized around three linked keyframe types and supports a workflow that moves from story planning to character and perspective exploration, and finally to narrative generation. Users: (1) create a story outline organized by plot keyframes; (2) define character keyframes to represent how each character develops across the story; (3) generate perspective keyframes to explore how those traits may be expressed in narrative form; (4) select the character traits and textual evidence they want to emphasize; and (5) generate third-person narratives conditioned on those keyframes and selections.

The interface consists of three coordinated views:

• Track view supports the main end-to-end workflow.

• Table view presents the outline, each character’s first-person perspective, and the generated third-person narrative side by side for each plot.

• Canvas view supports broader exploration by allowing writers to create multiple character arcs for a single character and combine different arcs across characters.

The design is grounded in narratological theories of character arcs, characterization, and focalization. Three design goals are derived: (DG1) shaping character development across plot points, (DG2) manifesting abstract character traits into concrete narrative evidence, and (DG3) tracing how characterization decisions propagate into generated story text.

The system was evaluated through two studies:

• Study 1 (Technical Evaluation): Using 20 writing prompts, stories generated by the system were preferred in 72 out of 100 pairwise comparisons by an automatic evaluator, with a significant difference in favor of the system (M = 6.24 vs. 5.74, t = 6.52, p <.001). Human raters preferred the system’s stories for richer characterization in 96.7% of pairs and for overall quality in 83.3% of pairs.

• Study 2 (User Evaluation): A within-subjects study with 12 writers found that narrative keyframing provided a significantly more controllable, transparent, and engaging creative experience than a baseline system. Participants rated the system significantly higher on transparency (M = 6.42 vs. 4.00, p =.003), controllability (M = 6.42 vs. 5.33, p =.047), enjoyment (M = 6.92 vs. 5.75, p =.010), immersion (M = 6.25 vs. 4.67, p =.011), exploration (M = 6.58 vs. 4.58, p =.004), and expressiveness (M = 6.58 vs. 5.08, p =.004).

The paper concludes that narrative keyframing supports a more controllable, transparent, and engaging way to use generative AI in creative writing, producing stories with higher overall quality and richer characterization than a standard LLM baseline. The work shows how keyframing can serve as an interaction paradigm for human-AI co-writing by helping writers balance automation with creative control.

Improvements to AI systems:

1. Implement a multi-layered narrative constraint engine that accepts structured keyframes (plot events, character trait trajectories, and perspective filters) as explicit input parameters, rather than relying solely on free-text prompts. This allows the AI to condition generation on discrete, temporally anchored story beats.

2. Add a character-state tracking module that maintains a dynamic vector representation of each character’s traits, emotions, and relationships at every story point. The AI updates this state after each generated segment, ensuring that prose consistently reflects the character’s defined arc (e.g., a shift from timid to assertive) across the narrative.

3. Introduce a perspective-switching mechanism that generates first-person internal monologues for any character at any plot keyframe, then uses those monologues as contextual anchors to produce third-person narration with consistent focalization. This enables the AI to control whose subjective experience colors the description of events.

4. Develop a keyframe-to-text alignment layer that maps high-level constraints (e.g., “character becomes more distrustful”) to concrete textual evidence (e.g., specific actions, dialogue, or internal thoughts) during generation. The AI can then explicitly tag which generated sentences satisfy which keyframe, improving transparency and editability.

5. Create a multi-view generation interface where the AI can output the same story segment in three coordinated formats: a track timeline (plot events), a table (character perspectives side-by-side), and a canvas (alternative character arcs). This allows the AI to support iterative refinement by regenerating only the prose between user-modified keyframes, while preserving the rest of the narrative structure.

6. Implement a controllability-aware reward model during fine-tuning that penalizes generated text that deviates from specified keyframe constraints (e.g., character acts out-of-trait or plot event is skipped). This improves adherence to user-defined narrative boundaries without sacrificing fluency.

What the improved AI system can do:

• Accept a story outline with plot keyframes, character development arcs, and perspective assignments, then generate coherent prose that strictly follows those constraints at each narrative moment.

• Automatically generate first-person perspective keyframes for any character, allowing the writer to preview how a scene looks from that character’s eyes before committing to third-person narration.

• Maintain a live, editable character state model that updates as the story progresses, so the writer can see how a trait change at one point propagates through later generated text.

• Offer a side-by-side comparison view of the same plot event written from different characters’ perspectives, with the AI explicitly highlighting which textual details are driven by each character’s traits.

• Support non-linear editing: the writer can move, add, or delete keyframes at any point, and the AI regenerates only the affected intervening prose while keeping all other segments consistent.

• Provide a “trait-to-evidence” trace feature, where the writer clicks on a character trait (e.g., “cautious”) and the AI highlights all generated sentences that instantiate that trait, making the creative process more transparent and revisable.

We introduce narrative keyframing, an interaction technique for AI-assisted creative writing that lets writers specify different types of narrative constraints at selected moments in a story, then use AI to generate intervening prose. Inspired by the use of keyframing in animation, narrative keyframing offers a flexible way to connect story planning with adaptive control over generated text. We explore three types of keyframes: plot keyframes define significant events in a story, character keyframes represent how individual characters change over the narrative, and perspective keyframes capture how individual characters experience different events through first-person narratives. Plot and character keyframes offer a flexible way to adapt the type of high-level conditioning explored in previous AI writing tools to more customizable, iterative, and fine-scale control, while perspective keyframes add a new way to control characterization and focalization by using first-person narratives as an intermediary. Through a user study, we show that narrative keyframing supports a more controllable, transparent, and engaging way to use generative AI in creative writing.