Knowledge Visualization: A Benchmark and Method for Knowledge-Intensive Text-to-Image Generation

Abstract: Recent text-to-image (T2I) models have demonstrated impressive capabilities in photorealistic synthesis and instruction following. However, their reliability in knowledge-intensive settings remains largely unexplored. Unlike natural image generation, knowledge visualization requires not only semantic alignment but also strict adherence to domain knowledge, structural constraints, and symbolic conventions, exposing a critical gap between visual plausibility and scientific correctness. To systematically study this problem, we introduce KVBench, a curriculum-grounded benchmark for evaluating knowledge-intensive T2I generation. KVBench covers six senior high-school subjects: Biology, Chemistry, Geography, History, Mathematics, and Physics. The benchmark consists of 1,800 expert-curated prompts derived from over 30 authoritative textbooks. Using this benchmark, we evaluate 14 state-of-the-art open- and closed-source models, revealing substantial deficiencies in logical reasoning, symbolic precision, and multilingual robustness, with open-source models consistently underperforming proprietary systems. To address these limitations, we further propose KE-Check, a two-stage framework that improves scientific fidelity via (1) Knowledge Elaboration for structured prompt enrichment, and (2) Checklist-Guided Refinement for explicit constraint enforcement through violation identification and constraint-guided editing. KE-Check effectively mitigates scientific hallucinations, narrowing the performance gap between open-source and leading closed-source models. Data and codes are publicly available at https://github.com/zhaoran66/KVBench.

• The paper introduces KVBench, a curriculum-grounded benchmark with 1,800 bilingual prompts spanning six academic subjects for objective scientific evaluation.
• The paper presents KE-Check, a two-stage T2I framework combining Knowledge Elaboration and Checklist-Guided Refinement to enforce structural and symbolic accuracy.
• The experimental analysis reveals that while closed-source models excel in reasoning and symbolic domains, multilingual robustness remains a key challenge.

Knowledge Visualization: Benchmarking and Methodology for Scientific Text-to-Image Generation

Motivation and Challenges in Knowledge Visualization

Text-to-image (T2I) generation has achieved remarkable visual realism and task adherence, but current models demonstrate substantial deficiencies when deployed in knowledge-intensive domains requiring strict compliance with scientific, structural, and symbolic constraints. Knowledge visualization diverges fundamentally from natural image synthesis, demanding granular reasoning and high-fidelity domain representation. Existing benchmarks and models often focus on subjective semantic plausibility, neglecting verifiable scientific correctness and structured constraint satisfaction.

KVBench: Curriculum-Grounded Benchmark for Knowledge Visualization

KVBench is established as a rigorous, curriculum-aligned benchmark aiming to bridge the assessment gap in knowledge-intensive T2I generation. It encompasses six high-school academic subjects—Biology, Chemistry, Geography, History, Mathematics, and Physics—with 1,800 bilingual prompts derived from over 30 authoritative textbooks. Each sample includes a reference image and a checklist of atomic evaluation criteria curated by subject-matter experts, supporting multi-dimensional, objective, and reproducible evaluation of structural and reasoning accuracy.

Figure 1: Illustrative examples from KVBench, featuring expert-curated prompts, textbook-grounded reference images, and atomic checklist-based evaluation criteria.

Data and evaluation pipelines are explicitly designed to ensure scientific accuracy and educational relevance. Prompts are dual-format: Brief Caption (high-level) for testing reasoning from implicit knowledge, and Detailed Caption (fine-grained) for explicit constraint rendering. Checklist-based verification replaces holistic scoring, decomposing visual correctness into binary, interpretable facts over entities, spatial relations, labels, and conventions; evaluated by high-performance MLLMs such as Qwen2.5-VL-32B.

Figure 2: Data construction pipeline (left) and evaluation pipeline (right) of KVBench.

KVBench covers a broad spectrum of knowledge visualization tasks and enables systematic diagnosis of model failures in reasoning, entity coverage, symbolic precision, and multilingual capability.

Figure 3: Examples from KVBench span diverse scientific and humanities knowledge visualization tasks across six academic disciplines.

KE-Check: Two-Stage Constraint-Aware Generation Framework

To address persistent deficiencies in knowledge-grounded generation, KE-Check is proposed as a modular, two-stage pipeline:

• Knowledge Elaboration: Input prompts are enriched via domain-specific expansion, integrating concept hierarchies, spatial arrangements, and visual conventions to generate structured descriptions.
• Checklist-Guided Refinement: Constraint satisfaction is enforced by constructing structured checklists, performing item-wise auditing to identify violations, and applying targeted, constraint-guided editing that preserves compliant content.

  Figure 4: KE-Check expands prompts and applies explicit checklist-based refinement for scientific fidelity.

This explicit decoupling mitigates scientific hallucinations and enhances structural, symbolic, and logical correctness in generated images. KE-Check demonstrates marked improvement in entity coverage, semantic logic, and visual readability.

Experimental Analysis: Model Performance and Error Taxonomy

Comprehensive evaluation of 14 SOTA T2I and multimodal models reveals:

• Closed-source models outperform open-weight counterparts in both reasoning-heavy and symbolic domains (e.g., Mathematics, Physics).
• Multilingual robustness is limited: Models consistently achieve higher scores under English prompts; cross-lingual discrepancies remain a challenge.
• Discipline-level disparities: Biology and Chemistry yield high fidelity due to standardized representations; History and Geography suffer from nuanced contextual demands.
• Prompt granularity matters: Detailed Caption yields improved performance in most cases, but can degrade output for weaker models due to instruction-following limitations.

  Figure 5: Performance comparison between prompt settings; Detailed Caption improves model performance in most cases, but impairs output for weak instruction followers.

Systematic error analysis classifies failures into three types:

• Visual Readability Failure: Garbled or unreadable text, predominantly afflicting open-weight models.
• Entity Coverage Failure: Missing or redundant entities, the most common error across all models.
• Semantic Logic Failure: Incorrect relationships or structure, especially prevalent in open-weight models, reflecting superficial pattern matching versus structured causal reasoning.

Ablation Results and Human Alignment

KE-Check's components are individually validated. Adding Knowledge Elaboration and refinement modules result in substantial gains (Baseline: 20.09/25.74; +Elaboration: 36.17/28.12; +Refinement w/o Checklist: 46.50/30.69; KE-Check: 47.67/30.24 for Chinese/English, respectively). Model evaluation protocol achieves strong human-AI alignment with overall accuracy of 80.26% and Cohen’s $\kappa=0.7447$, confirming the reliability of checklist-based automated judgment.

Qualitative Comparisons and Visualization

Qualitative visualizations across academic disciplines and languages demonstrate that KE-Check substantially reduces hallucinations and generates high-fidelity scientific diagrams.

Figure 6: Visualization of generated images on Detailed Caption in English; KE-Check yields structurally accurate textbook-quality outputs.

Practical, Theoretical Implications and Future Directions

KVBench establishes a foundation for rigorous, reproducible benchmarking of knowledge-intensive T2I models. The checklist-based granular evaluation protocol is directly translatable to pedagogy, scientific communication, and domain-specific illustration tasks. KE-Check’s modular approach sets a precedent for constraint-aware generative pipelines, clarifying the necessity of explicit knowledge grounding and post-hoc error correction. Future trajectories may include:

• Expansion of benchmark coverage to higher education and professional domains.
• Extension of checklist-guided editing via advanced programmatic compositional integration.
• Improvement of cross-lingual and symbolic rendering through task-specific training and architectural enhancements.
• Further research into robust multimodal evaluation metrics with fine-grained interpretability.

Conclusion

This work systematically addresses the gap in scientific and knowledge-intensive T2I generation by introducing KVBench—a curriculum-aligned, checklist-based benchmark—and KE-Check, an effective two-stage generation/refinement pipeline. Extensive empirical analysis reveals persistent model deficiencies in reasoning, symbolic fidelity, and multilingual robustness. The methods advanced here provide actionable direction for improving knowledge-grounded content creation and facilitate rigorous evaluation across pedagogically relevant domains (2604.22302).