TL;DR
While Vision-Language Models (VLMs) and text-to-image generators are becoming proficient at following explicit structural and safety rules for children's literature, they suffer from a significant "cultural alignment gap." This research demonstrates that AI systems consistently miss or misrepresent implicit cultural nuances—such as the difference between a Western shower and an Indonesian bathing dipper—failing to meet the standards of professional human illustrators.
What Problem Does This Paper Address?
Children’s book illustrations are more than just art; they are pedagogical tools governed by strict developmental and cultural guidelines. In Indonesia, the Pedoman Perjenjangan Buku (Book Leveling Guideline) codifies these requirements across different reading levels.
As AI enters the publishing workflow—both as a judge of content and a creator of it—a critical question arises: Can these models respect the "implicit" knowledge that human illustrators possess? Most AI models are trained on Western-centric datasets, leading to a potential erasure or caricature of non-Western cultures. The authors investigate whether AI can handle three pillars of evaluation: Guideline Compliance (explicit rules), Safety (age-appropriateness), and Cultural Alignment (implicit social norms).
Methods
The researchers conducted a two-task study:
- Task 1: VLM as Judge. They evaluated how well three VLMs—Claude Sonnet 4.6, Qwen3-VL 8B, and SEA-LION-v4-8B-VL (fine-tuned for Southeast Asia)—could categorize and assess 100 real Indonesian book pages compared to human experts.
- Task 2: AI as Generator. They used a "round-trip" prompting strategy to generate new illustrations. Claude Sonnet 4.6 first described an original page; this description was then combined with explicit cultural and guideline instructions to prompt two image generators: Imagen 4 and FLUX.1 [dev]. These images were then cross-checked by professional human illustrators.
Datasets
- Indonesian Ministry of Education Digital Library: 100 pages sampled across four early-literacy levels (A, B1, B2, B3).
- Human Baseline: Evaluations from three professional annotators for Task 1 and three professional illustrators for Task 2.
Key Insights and Ideas
- The "Shower vs. Dipper" Litmus Test: A primary example of the cultural gap is the bathroom scene. AI generators consistently depicted showers (common in the West), whereas Indonesian homes typically use a dipper and a standing tub (bak mandi). This subtle detail, captured by human illustrators, was missed by all AI models.
- Cultural Absence vs. Cultural Richness: The study found that while VLMs can detect the absence of cultural markers, they fail to recognize cultural richness. Human raters identified high cultural alignment in 23.3% of real pages, while the open-weight models (Qwen and SEA-LION) rated 0% of those same pages as having high cultural alignment.
- The "Batik Overload" Problem: When AI models do try to include cultural markers, they often over-apply them in unrealistic ways. For example, Imagen 4 placed identical batik patterns on mats, clothes, and toys simultaneously—a "cultural overemphasis" that feels unnatural to a local eye.
- Safety Sensitivity: Interestingly, Claude Sonnet 4.6 showed high safety sensitivity, even flagging a "false positive" in a published book where an adult knocking on a window appeared too aggressive for young children.
Key Results
- Level Classification: VLMs struggle with the nuance of reading levels. While they can distinguish between the simplest (Level A) and most complex (Level B3) books, they frequently collapse the middle levels (B1 and B2) into a single category.
- Human Preference: In Task 2, professional illustrators were unimpressed with AI output. When comparing Imagen 4 and FLUX.1, the most frequent verdict was "Both Bad" (42.2%), indicating that neither model met professional standards for children's publishing.
- Anatomy and Logic: Beyond culture, AI generators still struggle with basic physical logic in this domain, frequently producing "anatomy failures" (incorrect limb counts or distorted hands) that would be unacceptable in a children's book.
Limitations and Future Work
The authors acknowledge that their human-annotated subset was relatively small (N=20 for Task 1, N=12 pairs for Task 2) due to the high cost of expert labor. Furthermore, the study focused only on the four earliest literacy levels.
Future work should explore whether larger-scale fine-tuning on culturally specific datasets can bridge the "implicit knowledge" gap or if the problem requires a fundamental shift in how VLMs represent social and cultural context beyond simple object recognition.
Summary updated