💬 NLP

Narrative Keyframing for Generative Creative Writing

This paper introduces "narrative keyframing," an AI-assisted creative writing technique inspired by animation that enables writers to specify plot, character, and perspective constraints at specific story moments to generate intervening prose, thereby offering a more controllable, transparent, and engaging workflow for iterative storytelling.

Original authors: Chao Zhang, Abe Davis

Published 2026-08-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chao Zhang, Abe Davis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to tell a story. You might think the best way is to give the robot a giant list of rules: "The hero must be brave," "The villain must be sneaky," and "The story must have a happy ending." But here's the problem: robots are like over-eager interns. If you just give them a list of rules at the very beginning, they often forget the details by the time they get to the middle of the story, or they might make the hero act like a villain in the second chapter. This is a big puzzle for scientists who study how humans and computers work together (a field called Human-Computer Interaction). They want to know: How can we help a computer write a story that feels real, where characters actually grow and change, without the writer having to rewrite the whole thing from scratch every time?

To solve this, the researchers in this paper borrow a trick from the world of animation. In cartoons, animators don't draw every single frame of a character's movement. Instead, they draw a few "keyframes"—the most important moments where the character's pose or emotion changes drastically. The computer then fills in the smooth movement between those key moments. This paper asks: Can we do the same thing for writing? Can we give the AI a few "key moments" for a story's plot and character feelings, and let it fill in the rest? The authors, Chao Zhang and Abe Davis from Cornell University, built a new tool to test this idea. They call it "Narrative Keyframing." Their findings suggest that when writers use these special "key moments" to guide the AI, the resulting stories have much richer characters and feel more controlled than when writers just type a simple prompt and hope for the best.

The Animation Trick for Writers

Think of writing a story like directing a movie. In the old days, or when using standard AI tools, a writer might say, "Write a story about a girl and a wolf." The AI then tries to guess everything: what the girl thinks, what the wolf thinks, how they change, and how the story ends. It's like asking a director to improvise an entire movie without a script. The result can be okay, but the characters might feel flat or inconsistent.

The authors of this paper realized that professional writers often work differently. They don't just start writing; they plan the "key moments." They decide exactly what happens at the beginning, the middle, and the end. They also decide how the characters feel at those specific points. For example, a character might be "hopeful" at the start, "conflicted" in the middle, and "resolved" at the end.

The new tool, Narrative Keyframing, lets writers do this with AI. Instead of just typing a prompt, the writer uses three special types of "keyframes" to guide the computer:

Plot Keyframes: These are the big events. "The girl meets the wolf in the forest." "The girl gets lost." These are the anchors of the story.
Character Keyframes: These are snapshots of how a character feels or acts at those specific events. At the first meeting, the girl might be "curious but inexperienced." By the end, she might be "scared but brave." The writer sets these snapshots, and the AI figures out how the character gets from one state to the next.
Perspective Keyframes: This is the coolest part. The tool asks the AI to write the story from the character's own eyes (first-person) for each key moment. So, the AI writes a short paragraph as if it is the girl, saying, "He seemed kind, so I answered him honestly." Then, it writes another paragraph as if it is the wolf, saying, "She trusted me quickly, and I noticed it at once."

How the Magic Happens

Once the writer has these "first-person" snapshots, they act like a filter. The writer can look at what the AI wrote from the girl's perspective and the wolf's perspective and say, "I like how the girl sounded curious here, but I want the wolf to sound more sneaky there." The tool highlights the specific words and feelings the writer likes.

Then, the magic happens: the AI takes those selected pieces of "first-person" feeling and blends them together to write the final story in the third person (like "The girl looked at the wolf..."). Because the AI is pulling from the specific "keyframes" the writer chose, the final story feels much more consistent. The characters don't just randomly change their minds; they evolve exactly how the writer planned.

What the Tests Showed

The researchers didn't just build this tool; they tested it to see if it actually worked. They ran two main experiments.

First, they had the computer write stories using their new "Keyframing" method and compared them to stories written by the same computer using a standard "just give me a prompt" method. They asked human experts to read both sets of stories. The experts overwhelmingly preferred the stories made with the new tool. They said the characters felt more real, with specific details like "fingers trembling" or "knees itching," rather than just being described as "scared." The new method produced stories that were rated higher for both overall quality and for how well the characters were developed.

Second, they invited 12 real writers (from beginners to professionals) to try the tool. They asked the writers to write stories using the new tool and then using a standard chatbot tool. The results were clear:

More Control: The writers felt they had much more control over the story. Instead of guessing what the AI would do, they could set the character's mood at specific points and see it happen.
Better Understanding: The writers felt they understood the story better. Because the tool showed them the "first-person" drafts and highlighted which parts were used in the final story, they could see exactly how their choices turned into the final text. It wasn't a "black box" where magic happened; they could trace the path from their idea to the final sentence.
More Fun: The writers reported having more fun and feeling more immersed in the creative process. They felt like they were directing the AI rather than just asking it for help.

The Takeaway

The paper suggests that the way we talk to AI to write stories needs to change. We shouldn't just give it a big, vague instruction at the start. Instead, we should treat writing like animation: set the key moments, define how the characters feel at those moments, and let the AI fill in the gaps. By using these "keyframes," writers can keep their creative vision intact while still letting the computer do the heavy lifting of writing the prose. The tool doesn't replace the writer; it gives the writer a better way to steer the ship, ensuring that the characters grow and change exactly the way the story needs them to.

Technical Summary: Narrative Keyframing for Generative Creative Writing

Problem Statement

Recent AI writing systems have leveraged high-level story descriptions, character specifications, or plot outlines to condition the generation of narrative prose. However, these systems typically treat such conditioning as static prompts or global conditions. This approach limits fine-grained control over individual story elements and fails to capture how characters and plot points evolve dynamically throughout a narrative. Writers often struggle to balance automation with creative control, particularly when trying to manage character arcs (how a character changes over time) and focalization (how different characters perceive the same event) within a generated third-person story.

Methodology

The authors introduce Narrative Keyframing, an interaction technique inspired by animation keyframing. In animation, artists specify constraints at sparse key moments, and the system interpolates the states between them. The authors adapt this principle to generative creative writing by defining three linked types of narrative keyframes that connect story planning to narrative generation:

Plot Keyframes: These represent significant events in a story outline, serving as structural anchors for the narrative timeline.
Character Keyframes: These capture the state of a specific character at a particular plot point. They are organized into three dimensions based on Egri's character "bone structure": physiology (appearance), psychology (personality, values), and sociology (status, relationships). The system can automatically interpolate character states between defined keyframes to suggest transitions.
Perspective Keyframes: These are first-person narratives generated for specific characters at specific plot points. They serve as an intermediate representation to externalize abstract character traits into concrete textual evidence (actions, dialogue, thoughts, sensory details) based on Rimmon-Kenan's typology of characterization.

System Workflow

The proposed system supports an iterative workflow:

Planning: Writers define plot keyframes to outline the story structure.
Character Development: Writers define character keyframes at specific plots. The system suggests traits and interpolates character arcs between these points.
Perspective Exploration: The system generates first-person perspective keyframes for each character at each plot. Writers can inspect these narratives to see how traits are manifested.
Evidence Selection: Writers select specific textual evidence from the perspective keyframes to emphasize in the final story. The system highlights these selections bidirectionally, allowing edits to perspective text to update character keyframes and vice versa.
Narrative Generation: A third-person narrative is generated by synthesizing the selected plot points and the chosen textual evidence from the perspective keyframes. The output is color-coded to show the provenance of specific character traits.

The system was implemented using Next.js, React Flow for visualization, and GPT-4.1 for generation tasks (trait suggestion, interpolation, perspective generation, and final narrative synthesis).

Key Contributions

Narrative Keyframing Interaction: A novel interaction paradigm that allows writers to specify constraints at selected moments in a story rather than relying solely on global prompts. This enables fine-grained control over character arcs and narrative progression.
Perspective Keyframes as Intermediate Representation: The introduction of first-person narratives as a design material to bridge abstract character definitions and final third-person prose. This allows writers to concretize traits and selectively recombine evidence from different character viewpoints.
Bidirectional Traceability: The system explicitly links character traits, perspective evidence, and final narrative text, allowing writers to trace how specific characterization decisions propagate into the generated story.
System Implementation: A functional prototype integrating plot, character, and perspective views (Track, Table, and Canvas) to support the keyframing workflow.

Results

The authors evaluated the system through two studies:

Study 1: Technical Evaluation

Method: Generated stories using the authors' system versus a vanilla LLM baseline (GPT-4.1 with outline-to-narrative prompting) across 20 writing prompts.
Metrics: Automatic evaluation using the WQRM-PRE model and human evaluation by three expert writers on 30 story pairs.
Findings:
- The system's stories were preferred in 72% of cases by the automatic model.
- For characterization, human raters preferred the system's stories in 96.7% of pairs (29 out of 30).
- For overall quality, the system's stories were preferred in 83.3% of pairs (25 out of 30) under majority vote; however, when pooling all individual judgments, the preference for overall quality was 73.3%.
- Experts noted that the system produced "specific, physical characterization" and "embodied" emotional changes rather than summarizing them. However, some raters noted that the added detail occasionally impacted pacing or conciseness compared to the baseline.

Study 2: User Study

Method: A within-subjects study with 12 writers (ranging from beginners to professionals) who completed writing tasks using the proposed system and a baseline (character chatbots and outline-based generation).
Findings:
- Controllability: Users reported significantly higher control over character development and the ability to shape character portrayals at specific plot points ( $p < .01$ ). The staged workflow allowed for direct editing of character arcs rather than relying on conversational prompting.
- Manifestation: Users found the system significantly better at translating abstract character ideas into concrete story details ( $p < .01$ ). First-person perspectives were valued as a "window" to inspect how the AI interpreted character traits.
- Traceability: Users rated the system significantly higher for tracing how specific traits were reflected in the final text ( $p < .01$ ). Visual color-coding and explicit mappings reduced the cognitive load of verifying character consistency.
- Experience: The system scored significantly higher on the Creativity Support Index (CSI) for exploration, expressiveness, and immersion, and on the AI System Experience survey for transparency and controllability.

Significance and Claims

The paper claims that narrative keyframing offers a more controllable, transparent, and engaging way to use generative AI in creative writing compared to current static prompting methods.

Shift in Characterization: The work suggests that characterization should be viewed not as a static profile defined upfront, but as an evolving process that develops horizontally across plots and vertically across perspectives.
Perspective as Material: It posits that first-person perspectives can function as reusable design materials, allowing writers to concretize abstract ideas before final generation.
Traceability: The authors argue that making the connection between writer inputs and generated outputs explicit is crucial for generative creative writing, supporting a more reflective and deliberate mode of authorship.

The authors remain modest regarding their claims, acknowledging that their system supports a structured workflow which may feel restrictive for improvisational writers, and noting that their evaluation focused on short, three-act stories. They suggest future work could explore non-linear plots and other narrative properties like tone and pacing.