TL;DR
VideoMaker is a new co-creation system that treats creative intent not as a static prompt, but as an evolving shared state between humans and AI. By using a "Seed Panel" to externalize confirmed ideas and domain-specific knowledge bases to guide translation, the system helps creators maintain narrative continuity and professional control across the complex, multi-stage workflow of video production.
What Problem Does This Paper Address?
Generative AI has moved beyond simple media synthesis into the realm of creative planning and production. However, video creation is a multi-stage, multi-representational process (moving from ideas to scripts to storyboards to final frames). Current AI tools often suffer from three major issues:
- Intent Drift: AI reinterprets the creator's original idea as it moves between representations (e.g., a script detail getting lost in video generation).
- Homogenization: AI tends to suggest generic, formulaic narrative patterns, stifling unique creative directions.
- Lack of Negotiation: It is often difficult for creators to specify exactly which parts of an idea must remain fixed and which parts the AI is free to explore.
The authors argue that creative intent should be viewed as a dynamic state that must be continuously expanded, translated, and negotiated throughout the human-AI collaboration.
Methods
The researchers followed a user-centered design process, starting with semi-structured interviews with 12 creators to establish three core design goals (DG):
- DG1: Support differentiated expansion of intent.
- DG2: Support progressive and faithful translation across stages.
- DG3: Support continuously negotiable boundaries between human and AI.
They developed VideoMaker, a system built on a six-stage workflow:
- Input Creative Idea
- Co-create the Script
- Co-create the Storyboard Script
- Define Visual Style and Characters
- Generate Images and Videos
- Editing and Integration
Key technical components include:
- The Seed Panel: A central interface where confirmed intents are externalized as "seeds" (editable units). This creates a visible boundary: seeds are preserved, while empty space is open for AI suggestions.
- Multi-style Script Library: A database of 15 story structures and 15 narrative styles to help AI suggest non-formulaic directions.
- Cinematic Language Knowledge Base: 39 entries (shot sizes, camera angles, lighting) that act as "scaffolding" to translate abstract creator intent into professional audiovisual descriptions.
- Shared AI Agent: A unified agent (powered by Qwen3.6-35B) that carries the project context across all six stages.
Datasets
The paper does not use standard machine learning training datasets. Instead, it introduces two internal knowledge resources:
- Multi-style Script Library: Contains 87 linked cases of canonical films (e.g., Green Book, Thelma & Louise) categorized by structure and style.
- Cinematic Language Knowledge Base: A curated set of 39 professional filmmaking terms with definitions, visual characteristics, and reference images.
For evaluation, the authors conducted a within-subjects study with 30 participants and a quality assessment by 3 experts who rated the resulting narrative shorts.
Key Insights and Ideas
- Intent as a Shared State: The most significant shift is moving away from "text-in, video-out" toward a model where the human and AI jointly maintain a "living document" of intent.
- The "Seed" Metaphor: By turning vague ideas into discrete, editable "seeds," the system makes the "black box" of AI context visible and actionable.
- Semantic Slicing: While breaking the process into stages (DG2) helps professionalize individual shots, it can lead to "semantic slicing"—where shots are technically excellent but lack global coherence in lighting or tone.
- Negotiable Boundaries: The study found that the boundary between human control and AI initiative shouldn't be a one-time setting. It needs to be stage-adaptive: AI should be more "wild" during ideation and more "conservative" during final generation.
Key Results
In a comparative study against a baseline "chat-and-generate" system, VideoMaker showed significant advantages:
- Creativity Support: Participants rated VideoMaker significantly higher for Exploration (M=5.93 vs 4.43) and Differentiation (M=5.83 vs 4.20).
- Narrative Translation: The system was more effective at turning natural language into structured scripts (M=5.87 vs 4.47).
- Intent Preservation: The Seed Panel successfully helped creators keep their core ideas intact (M=5.33 vs 3.90).
- User Preference: Over 70% of participants expressed a clear preference for VideoMaker for long-term use, citing the visibility and traceability of the creative process.
Limitations and Future Work
- Global Consistency: Despite the use of seeds, maintaining visual consistency (e.g., character appearance) across many shots remains a technical challenge due to the stochastic nature of current video models.
- Cognitive Overhead: The structured, six-stage workflow was noted as being "slightly complex" for simple, short tasks, suggesting a need for a more flexible hierarchy.
- Local vs. Global Edits: Participants wanted better "isolation" of local changes—editing shot B shouldn't accidentally trigger a re-generation of shot A.
- Future Directions: The authors suggest developing stage-adaptive AI initiative and integrating more direct multimodal editing tools (like local brushes or multi-track timelines) into the co-creation flow.
Summary updated