How can a sequence of AI-generated illustrations retain the same characters while the story changes? This is the problem explored in a JTE paper by Minh Tan Le (Lê Minh Tân) and Dinh Tan Loc Nguyen (Nguyễn Đình Tấn Lộc), introduced by RTIC under the StoryVisual research project.
This Journal entry is an editorial introduction to the published work, not a new experiment or a claim that its methods were invented entirely by RTIC.
The paper and its authors
| Item | Published information |
|---|---|
| Paper | Application of Comics Generation Techniques in the Visualization of Vietnamese Folk Tales |
| Authors, in published order | Minh Tan Le; Dinh Tan Loc Nguyen |
| Vietnamese names | Lê Minh Tân; Nguyễn Đình Tấn Lộc |
| Affiliation | Ho Chi Minh City University of Technology and Engineering, Vietnam |
| Venue | Journal of Technical Education Science (JTE) |
| Online First | 25 August 2026, listed under Articles in Press |
| DOI | 10.54644/jte.2026.2148 |
In its 26 August announcement, RTIC identifies Lê Minh Tân as a Faculty of Information Technology lecturer and club academic adviser, and Nguyễn Đình Tấn Lộc as a Data Engineering student and member of the club's communications team. These descriptions are attributed to that dated announcement, not a newly verified roster appointment.
Selecting context instead of keeping everything
The study builds on Story-Iter, an existing iterative story-visualization architecture. Its proposed addition is a contextual filter: select relevant historical references rather than always carrying forward every image or a fixed run of the latest frames.
The paper uses SigLIP for semantic relevance in this reference-selection process. The distinction matters: Story-Iter and SigLIP are inherited components; the contribution discussed here is the integration and evaluation of the filtering strategy for Vietnamese folktale visualization.
An intuitive example is a recurring character who disappears for several scenes. The newest frames may be less helpful than an older, relevant scene. Choosing context deliberately can reduce irrelevant influence, although it cannot guarantee correct anatomy, identity or cultural detail in every generated image.
Evidence, with its limits intact
The authors report 23 Vietnamese folktales and 447 story segments in VNFairyTales. Their setup includes a 24 GB NVIDIA RTX 3090, 640 × 640 outputs and a fixed seed of 42. Evaluation considers text–image alignment, inter-frame similarity and character-related consistency, alongside a limited user study.
The useful conclusion is conditional: filtering can help when the reference window is constrained, but it does not improve every metric at every context length. Using the full reference set removes much of the distinction from the baseline. StoryDiffusion remains stronger on the reported CLIP-T comparison, and some character-consistency settings do not improve.
The paper also discusses a persistent difficulty: a semantically relevant reference can itself contain a visual error, which may be carried into later frames. A small, culturally specific dataset and limited human evaluation do not establish general superiority across all storytelling applications.
A research artifact worth following
For RTIC, the publication is a concrete example of lecturer–student research around a Vietnamese-language and cultural problem. Readers can inspect the paper and its linked artifacts rather than relying on a social-media headline or a single attractive image.
- Read the authoritative JTE article and full text.
- VNFairyTales repository referenced by the paper.
- Experimental outputs referenced by the paper.
- Author identifier: Minh Tan Le.
- Author identifier: Dinh Tan Loc Nguyen.
Research credit belongs to the paper's named authors. This summary is prepared by RTIC Editorial Desk. The article cover is an original AI-generated editorial illustration, not a figure or experimental output from the paper. Methods and results above are reported from the publication, not independently rerun.