Skip to content

Dual Coding

Principle · All principles
Evidence · 34 claims (16 for, 12 mixed, 6 against) · 44 studies (17 causal, 14 review, 10 quant-synthesis, 1 associational, 1 qualitative, 1 theoretical), q1–q4 · 14 of 44 report an effect size · 25 claims rest on one study

Conditional relationship

Some ideas have a structure that a picture can show directly: where parts sit, what causes what, the order of a process, how categories nest, how two things compare. For a learner who cannot yet describe that structure from words alone (a novice on this content), pairing the words that explain the idea with a visual that depicts the same idea is expected to improve comprehension and recall (conceptual-understanding, immediate-performance) over the words alone. The visual can be supplied with the text (a diagram, map, flow chart or labelled picture), or produced by the learner (a drawing, a concept map, a completed organizer). The visual has to carry the idea; a decorative image does not count and can cost something.

Claims test the two halves separately against words alone, mostly on immediate tests. A meta-analysis of 39 experiments found text with graphics beat text alone on reading comprehension, from elementary school to adults, and Mayer's laboratory programme found words with corresponding graphics beat words alone on transfer tests in 11 of 11 experiments. Meta-analyses found learner drawing in STEM and learner-built concept maps beat comparison conditions, and one tablet experiment found drawing beat reading only on transfer and drawing tests, not on recognition. No claim in the wiki tests the sequence below, the dose, or the move from a supplied visual to a learner-made one, and the two direct tests of constructing against studying run the other way: provided organizers gave better transfer than learner-built ones in three experiments, and building a concept map while viewing a text did no better than rereading, and worse than practising recall, a week later. Support for drawing did not help where it was tested. The supplied half has recorded exceptions too: an instructional image added to biology texts changed no test score in two experiments, and in physics problem solving a graph raised correctness on one problem and lowered it on another, while making learners more confident. Whether gains last (delayed-retention) is a separate claim that the evidence here barely touches.

The relationship is conditional on five things: whether the content has a visual structure at all; whether the visual shows the relation the task needs (a graph that invites quick confirmation can mislead); the learner's prior knowledge of it (a supplied visual may be redundant for an advanced learner); whether a learner-made visual is checked against a correct one; and how much time construction takes compared with other study. It does not rest on Paivio's two-code mechanism (Dual Coding Theory describes it): one set of experiments found the picture-superiority effect in memory followed physical distinctiveness, not dual coding. The design advice stands on comparisons of materials and activities, whatever the mechanism.

How the converted siblings sit beside this page: Multimedia Learning holds the presented explanation with a picture, where the picture goes, duplicated words, decorative additions and segmenting; this page does not repeat that model. Multimodal Instruction holds the choice of modes beyond words and a picture, spoken against written words, and translation between representations. This page holds the wider study practice: a visual paired with the words of a reading or a lesson, and the learner's own visual of the same idea, across several sessions. Retrieval Practice holds the comparison with recall from memory, which this page uses in its last steps. Creating Visual Representations, not yet converted, describes learner-made visuals as a general activity.

Default design

A concrete default a designer can act on and revise. Steps 2, 4 and 9 rest on claims that compare against words alone or no drawing; the rest are untested proposals, mostly from the earlier page.

  1. Check that the idea has a visual form. Use a visual where the content is a structure, a spatial relation, a sequence, a process, a comparison or a hierarchy. Where it is a list of unrelated items or an argument with no spatial or causal shape, a visual imposes an arbitrary structure; skip it. Untested proposal; from the earlier page ("clarify relationships, timing, scale, or process").
  2. Supply one relevant visual for each key idea, labelled in the words of the text. A picture, pictorial diagram or flow diagram all serve; choose the one that shows the relation the words describe, and prefer a diagram or animation of the process to a photograph of the thing. Remove images that are there only to look good. Basis: text with graphics against text alone [+M] (39 experiments, reading comprehension, elementary to adult; picture types did not differ); words with corresponding graphics beat words alone on transfer [+M] (Mayer's lab programme on brakes, pumps, lightning and arithmetic, 11 of 11 experiments, median d = 1.39); congruent pictures with text beat text alone or pictures alone [+M] (a 1993 review reporting Levie and Lentz second-hand); sketching and animation gave the best understanding of a geoscience text, a photograph the worst [~M] (120 undergraduates, six conditions); decorative illustrations [-M] (no gain, and they weakened the benefit of instructional pictures for 7th and 8th graders with lower prior knowledge). Qualified by an instructional image that changed no test score [-M] (biology texts, 163 and 48 undergraduates, d = 0.01 and 0.06), so a visual whose job is not stated may do nothing. Placement and labelling are modelled on Multimedia Learning.
  3. Ask for the link, not just the look. On the first encounter, give two or three prompts per visual: point to the part of the diagram this sentence names; say in words what the arrow shows; what in the text does this part of the picture stand for? Untested proposal (the claim pages suggest, without a study, that a visual nobody connects to the words does less). The nearest evidence in the wiki is prompted self-explanation [+M] (g = .55 across 69 effects, while studying text, examples or problems, mostly immediate tests); none of its entries prompts learners to explain how a picture and its words relate, so this step is carried from it, not tested.
  4. Then have learners make the visual themselves, with a clear instruction and little added support. Say what to draw (which parts, which relations, with labels) and let learners draw it unaided; give a partly completed diagram or map (key nodes given, links to label) only to a learner who cannot start, or where time is short, and treat it as a time saver, not a learning aid. Basis: drawing in STEM [+M] (a meta-analysis read from its abstract: drawing beat not drawing and beat imaging, explaining and re-watching; scaffolds saved time but did not raise outcomes); tablet drawing beat reading on transfer [+M] (156 learners randomised to read only, draw unsupported, or draw with global or local support; ηp² = .05, unsupported and locally supported drawing beat reading, global support did not differ); concept mapping [+M] (constructing maps showed a larger benefit than studying them, each against its own comparison). Against adding support: in the same tablet study, local support lowered recognition below reading only and drawing did not help recognition [-M], and both supports raised rated difficulty [-M]; the support rows on Scaffolding carry the drawing-accuracy result. Qualified by learner-built against provided organizers [~M] (in three experiments with college readers, provided organizers gave better transfer than learner-built ones) and by a narrative review that calls the evidence on learner-generated drawing inconsistent [~M]. Earlier versions of this step faded from a partly completed diagram to a blank page; the recorded tests of drawing support do not favour that, so it is now the fallback, not the default.
  5. Check each learner-made visual against a correct one. Have learners compare their drawing or map with a model, mark what is missing or wrong, and fix it, before it is used for study. Untested proposal; the claim pages warn, without a study, that an unchecked wrong diagram can fix a misconception.
  6. Follow the visual with recall, not more viewing. Once a learner has seen or made the visual, have them redraw it or explain it from memory, without the text, then check. Basis: concept mapping against retrieval practice [~M]: with learning time matched, undergraduates who practised free recall beat those who built a concept map while viewing the text a week later, and mapping was not significantly better than rereading. That the pairing of a visual with recall helps is an untested proposal.
  7. Plan the arc across sessions. A proposed default for a unit of about four sessions, untested: session 1, the supplied visual with the text and the linking prompts (about 10 minutes); sessions 2 and 3, the learner draws the idea from the text with a clear instruction of what to show (a partly completed diagram only for a learner who cannot start), then compares with the model (10–15 minutes); from session 4, open with a 5-minute redraw from memory of an earlier visual and check it, before new material. In mathematics, move from a picture of the quantity to the symbols across the same sessions, saying how each part of the picture maps onto the symbol (concreteness fading [~M], a review about concrete materials, carried here to pictures).
  8. Move on when the learner can redraw it. A learner moves on when they can draw the visual from memory with every key relation correct and explain each link in words, on two separate days. Untested proposal; set the criterion locally and record it.
  9. Judge it on a task that needs the relation. Use open-ended or mixed items (explain, predict, sketch), and decide in advance whether the visual may be consulted at the test. Basis: in the graphics meta-analysis, graphics helped more on open-ended and mixed-format tests than on true/false ones (text with graphics [+M]); in the tablet drawing study, drawing helped on transfer but not on recognition [-M], so a recognition-only test can miss what drawing taught.

When a condition is not met: if learners already have the model (advanced), drop the supplied visuals and go straight to a redraw-from-memory check; if there is no time for construction, keep the supplied visual and the linking prompts and use recall rather than mapping; if no one can check learner-made visuals, supply the model answer with the task; if learners will judge for themselves when they understand, have them redraw or explain from memory rather than rate their understanding.

Fitting the design to a situation

The three facts that most change the decision:

  • Whether the content has a visual structure. A diagram helps where it shows a relation the words describe; for list-like or purely verbal content it adds load and no meaning. If unstated, ask: what is the one relation (where, what causes what, in what order, which is part of which) a learner must grasp, and can it be drawn?
  • The learner's prior knowledge of this content. A supplied visual is a support for a learner who cannot build the model alone; for one who can, it may add nothing. If unstated, ask: can these learners, before instruction, sketch or explain the structure unaided?
  • Time, and who checks learner-made visuals. Drawing and mapping are slow, and a wrong diagram studied as right does harm. If unstated, ask: how many minutes per idea are there, and who will compare a learner's drawing with a correct one?
If the brief says… Then change… Basis
Learners are novice on this content Supply a complete, labelled visual first, with linking prompts; then ask for a drawing of named parts and relations, keeping a partly completed diagram for a learner who cannot start Graphic organizers for novices [~M] (lower-ability learners benefited from maps, higher-ability did not, in small, low-powered subsets)
Learners are advanced on this content Drop most supplied visuals; ask for a redraw from memory or a learner-made summary diagram, checked against a model Graphic organizers for novices [~M] (the higher-ability subset showed no benefit; the authors say this cannot show maps fail them); the redraw is untested
Learners are child (primary age) Supply pictures and flow diagrams with the text; have children label or complete them rather than draw freely Supplied: text with graphics [+M] (effect held across grade levels). Drawing: drawing in STEM [~M] (benefit larger in middle and high school than in primary)
Learners are adolescent or adult, goal is a principle in science or mathematics Add learner drawing of the process or system, with a model to compare against Drawing in STEM [+M] (STEM only, abstract read)
Goal is verbal-association (vocabulary, names, facts), including second-language vocabulary Pair each new word with a picture and a short text gloss; for students with learning disabilities, teach a keyword-image mnemonic L2: text-plus-picture glosses [+W] (one study of 90 Iranian EFL learners, immediate recall, reported second-hand). Disabilities: mnemonic instruction [+M] (mnemonics broadly, not only pictorial ones)
Goal is a concept or hierarchy to be understood across a unit Have learners build and revise a concept map across sessions, with labelled links, starting from given nodes Concept mapping [+M] (STEM and non-STEM; constructing showed more benefit than studying, against different comparisons)
online-self-paced, no teacher present Build a model drawing to compare with and a self-check list into the material; offer a drawing template as optional, and if the platform adds drawing aids (highlighting where to draw, prompts per part), compare them with plain drawing before keeping them Drawing in STEM [+M] (online drawing benefited more from instructor-provided scaffolds than paper-based drawing); against: in one tablet experiment the two supports tested did not beat unsupported drawing, and local support lowered recognition [-M] and raised rated difficulty [-M]
Learners solve problems with a graph or diagram supplied (physics, mathematics, engineering) Before they answer, ask what the graph shows and how it bears on the question; ask them to check an answer against the problem statement, not only against the graph; review which problems the graph actually informs The same graph helped on one problem and hurt on another [~M] (36 engineering-physics students; correctness .25 to .70 on one problem, .38 to .10 on another; with the graph, less use of prior knowledge and more "this must be correct"); no overall effect of graphs on correctness [-M] (56% against 52%)
Learners decide for themselves when they understand (self-study, revision, an adaptive course using self-ratings) Replace "rate how well you understand" with a redraw or explain-from-memory check after an illustrated text With an image, undergraduates rated their ability to draw the process higher [~M] (d = 0.32) though test scores did not differ [-M]; image-group learners judged by recalling the picture, text-only learners by generating their own image [~M] (self-reports of 18 and 23 learners, biology and chemistry)
Little time, or the goal is recall a week later Keep the supplied visual; replace building a map while viewing with redrawing or explaining it from memory Concept mapping against retrieval [~M] (undergraduates, science text, time matched, one-week test)
Large class, one teacher Collect a sample of learner drawings each session, show one correct and one common error to the class, then have everyone compare their own with the model untested proposal
Assessment is high-stakes or verbal only Include items that ask for a diagram or a prediction from one; decide whether the visual may be consulted Verbal tests may underestimate knowledge [~W] (an argument in a review, not a tested result); text with graphics [+M] (larger effect on open-ended items)
The brief describes "visual learners" or asks for materials matched to learning styles Give the visual to every learner where the content has a visual structure; do not sort learners or build matched versions Learning-styles matching [-S] (systematic review; 5th graders showed no style × modality interaction)
The designer wants images for appeal or engagement Keep only images that show what the words describe; if an image must stay for branding, keep it away from the explanatory visual Decorative illustrations [-M] (no learning gain; a small comprehension cost for undergraduates, η² = .04)

Twelve of the fourteen rows rest on a claim, and most of those claims rest on one study or an abstract; each row names the population tested, and the child, L2 and disabilities rows carry findings to closely related but not identical goals. The large-class row is an untested proposal, the redraw half of the advanced row is untested, and the self-judgment row rests on undergraduates reading short science texts online.

Observation, state and explanation

What can be observed is a learner's work under stated conditions: which visual they had or made, how long it took, what support they had, what they drew or labelled, what they got wrong, and what they could say or draw later without the material. "Has a mental model", "made a referential connection" and "is a visual thinker" are inferred states. A neat drawing shows drawing skill as well as understanding; a learner's liking for illustrated pages does not show that the pictures helped.

Response Plausible explanations Observation that could change the interpretation
Labels the supplied diagram correctly but cannot explain the process in words The labels were matched by position, not understood; the diagram shows parts but not the relation; the test asks for words the learner lacks Ask what happens when one part changes, with the diagram present and then removed; allow an answer by drawing or pointing.
Learner-made drawing is detailed but leaves out the key relation Effort went to appearance; the learner does not see which relation matters; the instruction asked for a picture, not a model Ask for a second drawing limited to five labelled parts and the arrows between them; compare with a model and ask what is missing.
Does well with the visual in view, poorly a week later The visual carried the structure and was not learned; little recall practice; interference from later material Give a redraw-from-memory task at the delay, and record all practice in between.
Concept map has many nodes and few labelled links Mapping was treated as listing; the learner lacks the relations; the format was not taught Ask the learner to label three links with a verb and explain one aloud; give a partly completed map with links to label.
Spends most of the time drawing and learns no more than from reading Construction cost exceeded its benefit; low drawing skill; the content had no visual structure; an added drawing aid took attention from the content Compare with a supplied visual plus linking prompts for the next idea, with time held equal; if a drawing aid was used, try the next drawing without it (in one tablet study, a local aid produced less accurate drawings and lower recognition than none, local support worse on recognition [-M]).
Says they understand an illustrated text well, then cannot redraw or explain it The judgment was cued by remembering the picture, not by building the model; the picture raised confidence without raising learning; the test asks for a relation the picture did not show Ask for a redraw or explanation from memory before accepting the judgment; compare with the same learner's judgment after a text-only passage. Learners given an image rated their ability to draw the process higher [~M] and said they judged by recalling the picture [~M].
Answers a problem with a graph present quickly and confidently, and wrongly The graph suggested an answer that was confirmed without checking; the graph does not show the relation the problem turns on; prior knowledge was set aside Ask the learner to say what the graph shows and which statement in the problem it bears on; give the same problem without the graph. With a graph, physics students used less prior knowledge and more often said "this must be correct" [~M].

These rows are proposals, untested as diagnostics. A probe is itself practice and changes what it measures, so record each one and keep the explanations that remain open.

Evidence and qualifications

  • Adding relevant pictures or graphics to text improves learning from it, with a small pooled effect on reading comprehension, though whether dual coding explains the benefit is disputed [+M]: a random-effects meta-analysis of 39 experiments (2,103 participants, 1985–2018; synthesis-experimental) comparing reading text with graphics against text alone found g = 0.39, 95% CI [0.26, 0.51], with moderate heterogeneity and no significant publication bias. The effect held across grade levels, picture types did not differ, and graphics helped more on open-ended and mixed-format tests. The outcome is reading comprehension, and some included studies found diagrams reduced verbatim recall. A narrative review (abstract only) reaches the same direction. The same page records experiments finding the picture-superiority effect follows distinctiveness, not dual coding, which challenges the mechanism the claim's opening states, not the effect. The page's title is broader than its evidence (comprehension from reading).
  • Words with corresponding graphics beat words alone on transfer [+M]: Mayer's 2008 summary of his programme's controlled experiments (paper and computer lessons on brakes, pumps, generators, lightning and arithmetic, mostly college students) reports 11 of 11 comparisons favouring words with graphics on transfer tests, median d = 1.39. One source summarising its own laboratory work, immediate tests; it is the transfer evidence behind step 2. A 1993 review reports, second-hand, that congruent pictures with text beat text alone or pictures alone [+M] on reading comprehension, with no effect size.
  • Recorded cases against a supplied visual (both held on Multimedia Learning, not repeated in full here): an instructional image added to biology texts changed neither test scores nor the accuracy of learners' judgments [-M] (Jaeger and Fiorella, 163 and 48 undergraduates, d = 0.01 and 0.06), and graphs did not raise correctness on vector-calculus problems [-M] (Ögren et al., 36 engineering-physics students, 56% against 52%). In the second study the same graph helped on one problem and hurt on another [~M]: correctness rose from .25 to .70 where the graph supported the dominant solution and fell from .38 to .10 where it invited confirmation; coded think-aloud data showed more use of the graph and less use of prior knowledge. Small samples, no effect sizes; they bound "add a visual" to visuals that show the relation the task turns on.
  • Sketching and animation gave the best understanding of a geoscience text, a photograph the worst [~M]: 120 undergraduates randomly assigned to six conditions while reading about climate change; condition mattered (ηp² = .19), with learner sketching and animation highest and photographs lowest. Pairwise contrasts are not on the claim page. It supports choosing a visual that shows a process over a realistic image, and a sketching task.
  • Drawing improves STEM learning by a small-to-moderate amount over not drawing and over other active study strategies, in one meta-analysis [+M]: a conference meta-analysis of 35 STEM studies (160 effects, 3,635 participants), read from its abstract: drawing beat not drawing (g = 0.362) and other active strategies (g = 0.400); scaffolds saved time (g = 1.693) but did not improve outcomes; larger benefit in secondary than primary school; online drawing benefited more from scaffolds than paper drawing. The claim page's opening still says its evidence base "has not yet been populated", which its own entry contradicts.
  • Tablet drawing beat reading on transfer [+M], but not on recognition, where local support did worse than reading [-M]: Schmidgall, Scheiter and Eitel (2020) randomised 156 learners to read only, draw with no support, draw with global support or draw with local support. Unsupported and locally supported drawing beat reading on transfer (ηp² = .05); no drawing condition helped recognition, and local support scored below reading (ηp² = .07); both supports raised perceived difficulty [-M] (ηp² = .10). Immediate tests, one study. It is the reason step 4 now defaults to plain drawing with a clear instruction.
  • The evidence on learner-generated drawing is inconsistent [~M]: Van Meter and Garner's 2005 narrative review sorts the mixed results by hypotheses about the processing drawing requires; no pooled estimate. It is a reason to check learner drawings and to test drawing against a supplied visual locally.
  • Concept mapping improves learning over most comparison conditions by a moderate average amount, but not over time-matched retrieval practice [+M] for construction, [~M] against retrieval: across 142 effect sizes (11,814 participants, abstract only), learning with maps beat comparison conditions (g = 0.58), constructing (g = 0.72) more than studying (g = 0.43), in STEM and non-STEM domains; an earlier meta-analysis of 55 studies (Grade 4 to postsecondary) agreed in direction. But with time matched, 80 undergraduates who practised free recall after studying a science text beat those who built a concept map while viewing it a week later (d = 1.50), and mapping was not significantly better than rereading; a second experiment (120 students) replicated the retrieval advantage, even on a map-building test. The page's own Discussion still says no studies are catalogued, which is stale.
  • Learner-constructed graphic organizers are not shown to outperform provided ones: the one direct test, with college readers, favoured provided organizers on transfer [~M], the limiting claim: in three experiments, college readers given author-provided graphic organizers did better on transfer than those who built their own (d = 0.24, not significant; 0.43; 0.84), with no difference on retention; the meta-analytic contrast favouring construction sets each condition against its own controls. The claim page says its title is a hypothesis the direct evidence does not support. It is why step 4 is scaffolded and checked rather than "let learners make their own".
  • Graphic Organizers Support Novice Comprehension [~M]: concept maps were associated with better retention overall (g = 0.60 across 67 effect sizes), less when studying a preconstructed map instead of text, lists or outlines (g = 0.37); in a few studies split at the median, lower-ability or lower prior knowledge learners benefited (g = 0.40, k = 5) and higher-ability learners did not (g = −0.13), subsets the authors call low-powered. It supports, weakly, supplying visuals to novices and withdrawing them for others.
  • Decorative Illustrations Do Not Improve Learning [-M]: a seductive-details meta-analysis (abstract only), three experiments with 7th and 8th graders (decorative pictures neither helped nor harmed overall but weakened the benefit of instructional pictures, most for lower prior knowledge), and an e-learning study with undergraduates (a very small comprehension cost, η² = .04; in harder texts, harm only for students with low attentional inhibition). It limits "add a visual" to visuals that carry the idea.

Learners, tasks and outcomes differ across these claims (readers of a text, STEM drawing, concept maps against recall, problem solving with graphs), so do not rank them by effect size. None in the wiki tests a supplied-then-constructed sequence or a delayed outcome of the full design.

The evidence by question

Further evidence, not yet read against this model

Claims the earlier page cited or that bear on this page but have not been used in the model above.

Objective and learner-valued goal

The designer's objective is usually that the learner can explain, predict or use the structure the visual shows, unaided, at a stated horizon. Elicit separately what the learner wants: to get through the reading, to remember it for an exam, to understand how something in their own life works, to have notes they can revise from. Agreement and divergence are both observations. A learner who values neat notes may spend the time on drawing quality; one revising for a recall test may prefer to reread a diagram rather than redraw it, which is the weaker choice for that goal.

For example, a nursing student may want to understand why a patient's blood pressure drops, while the designer's objective is labelling the parts of the heart. A labelled diagram serves the objective; a learner-drawn flow diagram of what happens when one part fails, checked against a model, tests the capability the student came for. Record both.

What would revise this model?

The expectation should weaken if comparisons with comparable learners and aligned outcomes find no advantage of a relevant supplied visual over text alone, or if the advantage seen on immediate tests is absent at a delay. If constructing visuals continues to lose to provided ones on transfer, or to recall practice with time matched, step 4 should give way to supplied visuals followed by redrawing from memory. If a drawing support is found to beat unsupported drawing on transfer, it should return to step 4 as the default. If added graphs keep helping on some problems and hurting on others, the page needs a rule for telling the two apart, which no claim in the wiki yet gives. If prompts to link words and visual add nothing over the visual alone, drop step 3. If the benefit of visuals turns out to depend on distinctiveness rather than meaning, the advice to remove decorative images would need re-examining. Do not protect the model by calling every failure a poor diagram or low drawing skill after the fact.

Comprehension while the visual is in view, recall a week later, transfer to a new problem and use in the setting the learner values are separate claims. The present evidence establishes no number of visuals per lesson, no length of drawing activity and no criterion for moving from supplied to learner-made visuals.

Claims

Examples

Key Sources

Page metadata
Field Value
type principle
id dual-coding
aliases creating-visual-representations
title Dual Coding
description For a learner without a working model of a structure or process, pairing the words that explain it with a visual of the same idea, supplied or learner-made, is expected to improve comprehension over words alone; claims test each half against words alone, none tests the sequence, direct tests favour provided over learner-built organizers for transfer, and the drawing supports tested did not help.
status review
canonical True
generated by claude/unspecified on 2026-10-19
trust unverified — no one has confirmed this page's content
sources 3 — cromley-2020, fiorella-2016, ainsworth-2006

Frontmatter is the page's machine-readable record — see the Schema & Guide.