Abstract
Qualitative narrative inquiry relies on "member checking"—the validation of researcher interpretations by study participants. However, this process is often hindered by the "dual burden" of high researcher workload and participant cognitive fatigue. This paper introduces NAME (Narrative-based Actor Masking and Embedding), a generative AI framework designed to transform dense text transcripts into coherent, visually grounded storyboards. By leveraging controllable image generation, NAME reduces the time required for qualitative validation from 800 hours to just 2 hours, while improving the clarity and emotional accessibility of research findings.
Introduction: The Challenge of Member Checking
In qualitative social science, researchers collect and interpret life stories. To ensure accuracy, they perform member checking, presenting their findings back to participants. Traditionally, this involves participants reading through volumes of academic text, which can be particularly taxing for survivors of trauma or those with limited literacy. This creates a significant cognitive and emotional burden.
The lack of automated tools to generate consistent visual narratives has left researchers with a choice between grueling manual synthesis or potentially alienating text-heavy reports.
Methodology: The NAME Framework
The authors propose NAME, built upon the StoryGen architecture. The framework focuses on "controllable generation" to ensure character consistency and precise spatial placement across a storyboard.
Technical Innovations
- Masked Cross-Attention Layer: NAME uses spatial masks and a differentiable noise-editing process to learn character placement during training, ensuring actors appear exactly where intended.
- Augmentation Module: To maintain image quality under spatial constraints, this module fuses mask-based features with CLIP text embeddings, a "dual-coding" approach that preserves high fidelity.
- Three-Stage Training Scheme: The model is trained progressively—first learning global semantics, then character identity, and finally spatial control. This prevents the model from prioritizing layout over character consistency.
Data and Evaluation
The study utilizes a curated subset of the StorySalon dataset. By using Segment Anything Model 2 (SAM2) for precise masking, the researchers achieved state-of-the-art results using less than 1% of the original data, highlighting the importance of data quality in narrative tasks.
To evaluate storytelling quality, the authors introduced three new metrics:
- Character Visual Consistency (CVC): Identity stability across frames.
- Spatial Narrative Consistency (SNC): Logical placement within the scene.
- Character Form Consistency (CFC): Realistic body proportions.
Key Results
The implementation of NAME yielded dramatic improvements in both efficiency and participant experience.
- Efficiency: The end-to-end pipeline was reduced from 800 hours to 2 hours.
- Cognitive Load: Human trials showed that participant "Reading Effort" scores dropped from 2.20 (text-only) to 1.16 (NAME visuals).
- Visual Quality: The model improved the Frechet Inception Distance (FID) score from 195 to 152 and outperformed baselines like StoryDALLE in character consistency (3.12 vs 2.38).
Conclusion
NAME demonstrates that generative AI can serve as a powerful tool for cognitive offloading in social science. By transforming abstract text into concrete visual anchors, researchers can foster more empathetic and efficient interactions with participants. Future work will focus on integrating these tools into standard qualitative analysis software to streamline the text-to-storyboard workflow.
Summary updated