Instructional Graphic Designer (IGD)

• Instructional Graphic Designer (IGD) is a professional who converts instructional goals into coherent visual artifacts by integrating design principles with pedagogical models.
• The role combines explicit, structured methods with tacit expertise, ensuring that visual compositions reflect both technical accuracy and learner-centered contexts.
• AI-mediated systems and layered design workflows enhance IGD outputs by supporting editable, structured visuals that meet rigorous educational and aesthetic standards.

An Instructional Graphic Designer (IGD) can be understood as a practitioner who translates instructional goals, process, and content into visual communication artifacts such as slides, diagrams, infographics, posters, handouts, LMS graphics, and animated explainers. In the recent literature, the role is defined less by a single occupational label than by a recurring technical and pedagogical problem: how to organize text, images, graphics, and interaction cues so that information is visually coherent, pedagogically structured, editable, and reusable across contexts. Research on instructional design, graphic design intelligence, and AI-assisted authoring places the IGD at the intersection of pedagogical modeling, visual composition, multimodal reasoning, and iterative production workflows (Chimalakonda et al., 2018, Huang et al., 2023, Kwak et al., 13 Mar 2026).

1. Disciplinary position

The instructional side of the role is anchored in explicit design models. A patterns-based approach to educational technology describes instructional design through GoalPattern, ProcessPattern, and ContentPattern, while also classifying instructional design knowledge under Context, Goals, Process, Content, Evaluation, and Environment. In that formulation, ProcessPattern can be organized as Play–Act–Scene–Instruction, and ContentPattern as Facts–Cases–Rules–Models–Theories (Chimalakonda et al., 2018). This makes visual production inseparable from instructional structure: the visual artifact is expected to encode goals, stage process, and differentiate content types.

RIGID broadens that framing by placing instructional work across Analysis, Design, Implementation, and Evaluation, and across micro, meso, and macro levels. Within that framework, generative AI is treated as a mediating layer that operationalizes research-based knowledge but does not replace expert judgment (Kwak et al., 13 Mar 2026). This suggests that an IGD is not merely a compositor of assets or a downstream “production” specialist. The role sits inside a larger evidence-oriented design cycle in which visual decisions are tied to learner characteristics, institutional constraints, and learning-sciences knowledge.

2. Knowledge base and expert judgment

A defining feature of the role is that much of its expertise is tacit rather than fully codified. A literature review and interview study with 10 professional graphic designers identified 14 tacit knowledge characteristics across Possession, Expression, Acquisition, and Manifestation, and collected 123 tacit knowledge instances concentrated around inner design elements, cognition and manipulation actions, and visual and audience-oriented purposes (Son et al., 2024). The same study describes tacit design knowledge as knowledge that is difficult to communicate verbally or through text, is learned by doing and using, and is specific to context, professional experience, intuition, and convention.

The examples are close to the actual work of instructional graphics: “adjusting the right font size, boldness, line spacing, and margin according to the hierarchy of the content,” “creating a layout that guides the audience's actions,” and “predicting the mental model of the target audience” (Son et al., 2024). These are not reducible to fixed templates or isolated software commands. They are judgments about hierarchy, readability, pacing, emphasis, and audience response.

This suggests that the IGD’s knowledge base is hybrid. Part of it is explicit—design principles, structured authoring methods, layout conventions, Bloom-oriented instructional objectives. Part of it is embodied and situational: recognizing when a slide is technically correct but misprioritized, when spacing is visually uncomfortable, or when a visual metaphor will distract rather than explain. The literature on tacit knowledge is therefore directly relevant to the IGD because it explains why strong instructional visuals are difficult to automate or standardize fully.

3. Workflow, representations, and deliverables

Operationally, the role spans content preparation, asset generation, composition, and local refinement. One upstream bottleneck is content decomposition. Work on document chunking and learning-objective generation treats source documents as reusable instructional material, breaking them into chunks and attaching objective language that incorporates Bloom’s verbs (Tran et al., 2018). That approach does not produce graphics by itself, but it creates the structured instructional inputs that later visual design depends on.

At the authoring level, infographic creation has been decomposed into six essential task categories: Information Collection, Visual Element Design, Pivot Figure Design, Background Design, Layout Customization, and Local Adjustment (Huang et al., 2024). This decomposition is especially compatible with instructional graphics, where designers often begin from a topic or lesson idea and must derive title text, supporting points, icons, central imagery, background treatment, and final composition from scratch.

Recent systems increasingly externalize the design artifact as structure rather than as pixels. DocLap predicts CSS-like fields—top, left, width, height, and layer—from component images plus instructions about design purpose and canvas size (Zhu et al., 2024). Graphist formalizes Hierarchical Layout Generation as prediction of element geometry and hierarchy, representing each element by $(x_i, y_i, w_i, h_i, l_i)$, and emits a JSON draft protocol describing source identity, category, semantic type, coordinates, size, and order (Cheng et al., 2024). These representations suit IGD work because they preserve editability and make explicit the relation between instructional role and graphic role: title, image, accessory, text underlay, sticker, or background are not merely pixels but typed layers in a draft composition.

4. AI-mediated production paradigms

AI support for the IGD has moved from narrow layout tools toward multimodal design systems. In layout planning, DocLap trains on Coordinates Predicting, Layout Recovering, and Layout Planning, and on Crello reaches 43.75 mIoU versus 35.17 for 1-shot GPT-4V (Zhu et al., 2024). LayoutDETR reformulates multimodal layout generation as a detection problem over a background image, predicting reasonable locations, scales, and spatial relations for foreground text and image elements (Yu et al., 2022). CreatiDesign pushes this further with a unified multi-conditional diffusion transformer on FLUX.1-dev, combining a global prompt, a multi-subject image condition, and semantic layout tokens, with multimodal attention masks that bind each condition to its designated image region (Zhang et al., 25 May 2025).

A second line of work emphasizes layered editability. Accordion takes a top-down approach: it first obtains a full reference image and then decomposes it into object layers $\{O_1,\dots,O_N\}$, a background layer $B$, and a vectorized text layer $T$, while also refining nonsensical AI-generated text into meaningful alternatives guided by prompts (Chen et al., 8 Jul 2025). PSDesigner instead learns a human-like PSD workflow: it gathers theme-related assets, predicts Photoshop tool calls in asset-integration and layer-refinement modes, and outputs editable PSD files with complex layer hierarchy, masks, effects, and adjustment layers (Shuai et al., 26 Mar 2026). A related step is the model titled “IGD: Instructional Graphic Design with Multimodal Layer Generation”, which uses an MLLM for attribute prediction, sequencing, and layout of layers, while reserving diffusion for image-asset generation (Qu et al., 14 Jul 2025).

When the artifact is procedural rather than page-based, generation shifts from static composition to sequential illustration. StackedDiffusion generates illustrated instructions by pairing LLM-produced step text with jointly generated step images and evaluating them through goal faithfulness, step faithfulness, and cross-image consistency (Menon et al., 2023). $I^2G$ models a procedure as a goal plus ordered steps, uses constituency-parser-based text encoding, pairwise discourse coherence, and Stable Diffusion XL, and reports consistent gains over SDXL and other baselines on HTStep, CaptainCook4D, and WikiAll (Bi et al., 22 May 2025). For the IGD, these systems matter because they extend visual design from posters and slides to recipes, guided tasks, assembly-like procedures, and other instruction sequences in which continuity and intermediate state depiction are central.

5. Data, evaluation, and evidence

The research infrastructure around the role is increasingly structure-first. Two surveys organize graphic design AI around representation, understanding, and generation, or around perception and generation, and both argue that future systems must connect localized visual features with global design intent (Huang et al., 2023, Zou et al., 24 Mar 2025). LICA extends this direction with 1,550,244 multi-layer graphic design compositions, typed components including text, image, vector, and group elements, and 27,261 animated layouts with per-component keyframes and motion parameters (Hirsch et al., 17 Mar 2026). For the IGD, such datasets are important because instructional artifacts are typically layered, typed, editable, and frequently animated or repurposed.

Evaluation has also become more multidimensional. TASTE shows that designers judge AI-generated graphic design along distinct criteria including Mood and Tone Match, Visual Hierarchy, Color Harmony, Typography, Color Accuracy, and Spatial Accuracy, and that no pre-trained VLM judge or dedicated text-to-image scorer in its benchmark exceeds 0.55 macro agreement with the 5-designer majority; a small pairwise-difference head trained on TASTE reaches 0.611 against a 0.741 single-rater ceiling (Zhu et al., 20 May 2026). This is especially relevant to instructional graphics, where overall preference is a weak substitute for separate judgments about hierarchy, text fidelity, and spatial logic.

Task-specific work adds narrower but technically useful metrics. Graphist introduces Inverse Order Pair Ratio for overlapping layer order and GPT-4V-based ratings such as SDL, SGI, SIO, and STV (Cheng et al., 2024). Procedural illustration work supplements image-quality metrics with KL divergence, Chi-square, and caption-based long-text alignment so that a generated image is evaluated not only for realism but for whether it actually depicts the intended instructional step (Bi et al., 22 May 2025). The cumulative implication is that the IGD’s output cannot be evaluated by geometry or aesthetics alone; it demands structured assessment of readability, hierarchy, semantic alignment, and editability.

6. Limitations, controversies, and future directions

Across the literature, a common misconception is that stronger generation removes the need for human expertise. The opposite conclusion is stated repeatedly. RIGID treats AI as a mediating layer while preserving the central role of human expertise (Kwak et al., 13 Mar 2026). DocLap is explicitly better understood as a layout planning copilot than as an autonomous final designer (Zhu et al., 2024). GraphiMind couples a conversational interface with a graphical manipulation interface because local adjustment, styling, and iterative revision remain necessary parts of the workflow (Huang et al., 2024). The role of the IGD therefore shifts, but does not disappear.

The hardest unresolved problems are concentrated precisely where instructional graphics are most demanding. Reported limitations include typography quality beyond bounding-box placement, pedagogical clarity and sequencing, accessibility concerns such as contrast, reading order, and screen-reader structure, semantic correctness of diagrams and chart relationships, content accuracy, brand/style consistency across a curriculum, and classroom appropriateness (Cheng et al., 2024). Other systems still struggle with abstract or implicit actions, long-horizon consistency, dense text, facial detail preservation, and lexical failures such as the reported “steak” $\rightarrow$ “angel” error in procedural illustration (Bi et al., 22 May 2025, Zhang et al., 25 May 2025). Tool support is also weakest exactly where professional judgment is strongest: cognition actions, audience-oriented purposes, and content-related decisions remain under-supported relative to more mechanical workflow assistance (Son et al., 2024).

The strongest near-term direction is therefore not fully autonomous design but research-integrated, structure-first, human-in-the-loop systems: editable layered outputs, richer datasets with operation traces, criterion-specific evaluators, and authoring environments that can explain, revise, and preserve design rationale across series of related assets (Shuai et al., 26 Mar 2026, Hirsch et al., 17 Mar 2026, Zhu et al., 20 May 2026). This suggests that the IGD is increasingly defined by supervision of multimodal generation, validation of textual and pedagogical fidelity, and maintenance of reusable visual systems rather than by manual placement alone.