CAGE: Bridging the Accuracy-Aesthetics Gap in Educational Diagrams via Code-Anchored Generative Enhancement

Dikshant Kukreja; Kshitij Sah; Karan Goyal; Mukesh Mohania; Vikram Goyal

cs.CVApr 6, 2026

CAGE: Bridging the Accuracy-Aesthetics Gap in Educational Diagrams via Code-Anchored Generative Enhancement

Authors: Dikshant Kukreja, Kshitij Sah, Karan Goyal, Mukesh Mohania, Vikram Goyal

Organizations: IIIT Delhi, India

Abstract

Educational diagrams -- labeled illustrations of biological processes, chemical structures, physical systems, and mathematical concepts -- are essential cognitive tools in K-12 instruction. Yet no existing method can generate them both accurately and engagingly. Open-source diffusion models produce visually rich images but catastrophically garble text labels. Code-based generation via LLMs guarantees label correctness but yields visually flat outputs. Closed-source APIs partially bridge this gap but remain unreliable and prohibitively expensive at educational scale. We quantify this accuracy-aesthetics dilemma across all three paradigms on 400 K-12 diagram prompts, measuring both label fidelity and visual quality through complementary automated and human evaluation protocols. To resolve it, we propose CAGE (Code-Anchored Generative Enhancement): an LLM synthesizes executable code producing a structurally correct diagram, then a diffusion model conditioned on the programmatic output via ControlNet refines it into a visually polished graphic while preserving label fidelity. We also introduce EduDiagram-2K, a collection of 2,000 paired programmatic-stylized diagrams enabling this pipeline, and present proof-of-concept results and a research agenda for the multimedia community.

Figures & tables

Panel of Figure 1 — Figure 1. The accuracy–aesthetics dilemma visualized on the prompt “labeled diagram of photosynthesis.” (a) Open-source diffusion: produces engaging visuals with catastrophically garbled labels. (b) Code-based generation: guarantees label accuracy but yields visually flat output. (c) Closed-source APIs: improve fidelity at significant per-image cost but remain imperfect (e.g., “Stomaata”). (d) CAGE: achieves both accuracy and visual quality at zero marginal cost.

Scenario	DALL·E 3	GPT-4o	Gemini
Per image ($)	0.040	0.080	0.040
Per deck ($)	0.48	0.96	0.48
Teacher/yr a ($)	19.20	38.40	19.20
School/yr b ($)	960	1,920	960
Eff. cost c ($)	1.3–1.6 $\times$ above

Table 1. Cost analysis of closed-source diagram generation at educational scale. Assumes 12 diagrams per deck.

Paradigm	Model	LEM $\uparrow$	CER $\downarrow$	FID $\downarrow$	HVA $\uparrow$	$/img
Open-src. diffusion	SDXL 1.0	11.3	71.4	112.6	3.8	0
	Flux.1-dev	18.7	59.2	95.3	4.1	0
	SD3 Med.	14.9	64.8	103.7	3.9	0
Code-based (LLM $\to$ code)	GPT-4o	97.2	0.8	184.5	2.1	0 a
Code-based (LLM $\to$ code)	Claude 3.5	95.6	1.4	191.3	2.0	0 a
Closed-src. APIs	DALL·E 3	64.3	19.7	98.4	4.0	0.04

Table 2. Accuracy and visual quality across paradigms on 400 K–12 diagram prompts. LEM = Label Exact-Match (%); CER = Character Error Rate (%); FID = Fréchet Inception Distance (against textbook reference,

\downarrow

); HVA = Human Visual Appeal (1–5,

\uparrow

; Krippendorff’s

\alpha=0.78

n{=}6

annotators).

Panel of Figure 2 — Figure 2. The CAGE pipeline. Stage 1 synthesizes executable code via an LLM, executes it to produce a structurally correct programmatic rendering $I_{\text{prog}}$ , and verifies label completeness in a REPL-style loop. Stage 2 extracts Canny edges from $I_{\text{prog}}$ as a spatial “cage,” then applies ControlNet-conditioned SDXL diffusion to produce $I_{\text{ref}}$ —a visually polished output that preserves all labels and diagram topology.

Panel of Figure 3 — Figure 3. Sample pairs from EduDiagram-2K. Each row shows a code-generated diagram ( $I_{\text{prog}}$ , left) and its CAGE-refined stylized version ( $I_{\text{style}}$ , right). All labels and structural topology are preserved while visual quality is substantially enhanced.

Property	Value
Total pairs	$\sim$ 2,000
Subjects	Biology, Chemistry, Physics, Mathematics
Grade bands	K–5, 6–8, 9–12
Diagram types	Flowcharts, labeled illustrations, concept maps, structural diagrams
Avg. labels/diagram	8.4
Rendering languages	Python, L a T e X /TikZ, SVG

Table 3. EduDiagram-2K dataset statistics.

T2I Generation

Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.

See all 299 papers

EduIllustrate: Towards Scalable Automated Generation Of Multimodal Educational Content
Apr 6, 2026Shuzhen Bi, Mingzi Zhang, Zhuoxuan Li +3Generative AI in Education Multimodal Generation
Shanghai Innovation Institute · University of Science and Technology of China · East China Normal University
Math-Vision Diagrams: A Comprehensive Benchmark for Evaluating LLM Mathematical Diagram Generation Capabilities
Aug 9, 2026Harish Kashyap, Kiran Byadarhaly, Sriram Chakaravarthy +2LLM Evaluation T2I Generation Evaluation
Pandita AI Inc., 2540 N First St, San Jose CA 95131, USA
EvoDesign: Agentic Editable Diagram Creation via Design Expertise Evolution
Feb 20, 2026Tianfu Wang, Leilei Ding, Ziyang Tao +12Self-Evolving Agents
Hong Kong University of Science and Technology (Guangzhou) · University of Science and Technology of China · Peking University +1

Apr 6, 2026J/K move · Enter open · S save

EduIllustrate: Towards Scalable Automated Generation Of Multimodal Educational Content

Shuzhen Bi, Mingzi Zhang, Zhuoxuan Li, Xiaolong Wang+2

Shanghai Innovation Institute · University of Science and Technology of China · East China Normal University

Generative AI in Education Multimodal Generation Education

Open paper AskPDF arXiv

Large language models are increasingly used as educational assistants, yet evaluation of their educational capabilities remains concentrated on question-answering and tutoring tasks. A critical gap exists for multimedia instructional content generation -- the ability to produce coherent, diagram-rich explanations that combine geometrically accurate visuals with step-by-step reasoning. We present EduIllustrate, a benchmark for evaluating LLMs on interleaved text-diagram explanation generation for K-12 STEM problems. The benchmark comprises 230 problems spanning five subjects and three grade levels, a standardized generation protocol with sequential anchoring to enforce cross-diagram visual consistency, and an 8-dimension evaluation rubric grounded in multimedia learning theory covering both text and visual quality. Evaluation of ten LLMs reveals a wide performance spread: Gemini 3.0 Pro Preview leads at 87.8%, while Kimi-K2.5 achieves the best cost-efficiency (80.8% at \

0.12/problem). Workflow ablation confirms sequential anchoring improves Visual Consistency by 13\% at 94\% lower cost. Human evaluation with 20 expert raters validates LLM-as-judge reliability for objective dimensions (

ρ\geq 0.83$) while revealing limitations on subjective visual assessment.

Abstract

T2I Generation

Explore similar work

EduIllustrate: Towards Scalable Automated Generation Of Multimodal Educational Content

Math-Vision Diagrams: A Comprehensive Benchmark for Evaluating LLM Mathematical Diagram Generation Capabilities

EvoDesign: Agentic Editable Diagram Creation via Design Expertise Evolution