Internet Meme: Multimodal Cultural Artifact

• Meme is a multimodal cultural artifact that combines images, text, and context to create meaning through reuse and variation.
• Computational research applies template-based classification, multimodal similarity, and context-aware reasoning to decode meme semantics.
• Studies use diverse datasets and benchmarks to improve meme moderation, retrieval, generation, and interpretability in digital media.

A meme, in contemporary computational and memetics-oriented research, is not merely an image with text but a multimodal cultural artifact whose meaning arises from reuse, variation, and context. One operational definition describes a multimodal Internet meme as “a multimodal user-generated piece of digital content that contains at least one visual or textual component that can be identified as a memetic element that is shared with other related but distinct memes,” while a complementary definition treats an internet meme as a group of digital items that share common characteristics, are created with awareness of each other, and are circulated, imitated, and transformed via the Internet by multiple users (Hazman et al., 2024). In computational work, memes are therefore studied simultaneously as cultural lineages, template-based communicative forms, multimodal classifiers’ inputs, and increasingly as context-sensitive social acts whose interpretation depends on image, embedded text, surrounding discourse, and external knowledge (2505.17433).

1. Conceptual foundations

The memetics-grounded view distinguishes the meme itself from any single instantiation. In this view, a “memetic element” is a textual or visual element observably reused in multiple distinct memes, and “related but distinct” refers to meme pairs or groups that share such elements while differing by some novel component, such as different captions or inserted foreground material (Hazman et al., 2024). A closely related empirical formulation models a meme as the conjunction of an image template and a tacit concept that the template evokes; implementations are captioned realizations of that concept, and the meme can be written as $m=\{i_m^1,i_m^2,\ldots,i_m^n\}$ (Coscia, 2013). This distinction between template-level semantics and instance-level adaptation is central to later computational treatments.

The same conceptual line appears in studies of templatic memes. A template is treated as a recognizable base pattern whose underlying meaning is known within a community, while the individual meme instance adapts that meaning to a particular communicative intent through overlay text, cropping, substitution, or remixing (Bates et al., 2023). Context-oriented work broadens the definition further by treating memes as “amateur media artifacts, extensively remixed and recirculated by different participants on social media networks,” emphasizing that their meaning is not exhausted by the artifact alone but is negotiated in community use (2505.17433).

2. Structural forms and semantic mechanisms

A memetics-grounded typology identifies several recurrent forms. “Template Memes” include “Character Macro” memes, such as “Imagination Spongebob” and “First Day on the Internet Kid,” and “Format Macro” memes such as the two-panel “Drake” format. Other categories include “Memetic Images,” which reuse an image plus appended whitespace for changing captions; “Transferred Symbols,” in which localized symbols such as the “Disaster Girl” face or “Monkey Puppet” are inserted into new contexts; and “Memetic Trend,” where textual slogans such as “Better love story than Twilight” spread across visually dissimilar items (Hazman et al., 2024). This typology connects visual reuse, symbolic transfer, and linguistic convention under a single account of memetic inheritance and variation.

Semantically, memes are difficult because image and text do not simply add information; they frequently invert, refract, or metaphorize each other. Work on meme captioning and interpretation characterizes many memes as metaphorical: the image often serves as a vehicle for an abstract target, and the literal image content should not appear in the final interpretation. In MemeCap, 44% of a manually analyzed subset were “complementary,” meaning that interpretation required both modalities, while person or character vehicles and behavior- or stance-oriented targets were especially common (2305.13703). Other work reaches a similar conclusion through multimodal similarity and emotion analysis, arguing that image-only or text-only pipelines overlook the meaning that emerges from their interaction (Konyspay et al., 21 Mar 2025).

Context can change meme meaning even when the image-text artifact remains unchanged. MemeReaCon formalizes this with the distinction between “Context Explain Meme” and “Meme Enhance Context,” and illustrates it with a “Stack” versus “Heap” meme whose intent flips between a programmer community and a general-audience forum (2505.17433). A plausible implication is that meme semantics are often distributed across artifact, template lineage, discourse frame, and audience background rather than located in any one channel.

3. Datasets and benchmark ecosystems

Meme research is organized around a heterogeneous benchmark landscape that targets different aspects of memeticity, affect, harm, interpretation, retrieval, and generation. Representative resources include the following.

Resource	Focus	Scale
KYMKB	template knowledge base	5,220 base templates; 54,751 images
Memotion Dataset	humor, sarcasm, motivation, offensiveness, sentiment	6,992 meme images
MemeCap	meme captioning, literal captions, visual metaphors	6,384 memes
MCC	explanatory evidence for memes	3,400 English memes
ExHVV	hero, villain, victim explanations	3K memes; 4,680 entities
MemeReaCon	meme understanding in Reddit context	1,565 instances
CM50	templatic memes with automated annotations	33,173 memes
MER-Bench	meme reappraisal	3,117 real-world memes

These resources are complemented by corpora such as the Reddit Memes Dataset with 3,226 meme images and associated titles and timestamps, the Hateful Memes dataset with 10,000 images used for binary meme-vs-regular detection, the Meme Caption Dataset with 177,942 caption instances across 24 templates, and the 1.1 million-caption Memeify corpus spanning 128 classes (Konyspay et al., 21 Mar 2025, Koutlis et al., 2022, Sadasivam et al., 2020, Vyalla et al., 2019). Template-centric work has also introduced CM50, centered around 50 popular meme templates and enriched with image captions, meme captions, and literary device labels (Deng et al., 23 Jan 2025).

At larger scale, MemeLens consolidates 38 public meme datasets across nine languages and reports final totals after preprocessing of approximately 178K train, 22K validation, and 40K test samples, with task mappings spanning harm, targets, figurative or pragmatic intent, misinformation or persuasion, and affect (Shahroor et al., 18 Jan 2026). This suggests that the field has moved from isolated benchmarks toward cross-dataset and multilingual infrastructures, although the underlying label spaces remain highly heterogeneous.

4. Computational modeling, matching, and retrieval

A large part of computational meme research treats the meme as a multimodal classification or retrieval object. For image-macro detection, MemeTector introduces “Visual Part Utilization,” which pairs each meme with a text-free crop of its own background and trains a ViT-based classifier with an additional trainable attention module. Across 169 train/test scenario crossings, it reports accuracy between 89.0% and 97.8%, with mean 94.98% and standard deviation 1.47%, outperforming vanilla ViT in 148 of 169 scenarios (Koutlis et al., 2022). For broader meme classification, SimCLIP uses frozen or partially thawed CLIP encoders and Siamese fusion, establishing a new state of the art on Memotion7k with a 7.25% relative F1-score improvement and achieving 80.01% F1 on Harm-P, above the reported human F1 of 70.35% (Huertas-Tato et al., 2024).

Similarity-based work frequently operates directly in a shared image-text space. One such formulation groups memes by thresholded cosine similarity,

$s(x,y)=\frac{x\cdot y}{\|x\|_2\|y\|_2},$

using image-image, text-text, and cross-modal comparisons, while Siamese CLIP fusion explicitly encodes modality agreement and disagreement as

$$z=[e_{txt},e_{img},|e_{txt}-e_{img}|,e_{txt}\odot e_{img}].$$

The former is used for threshold-based grouping of similar memes, and the latter for compact cross-modal classification (Konyspay et al., 21 Mar 2025, Huertas-Tato et al., 2024).

Template-aware methods replace heavy end-to-end fine-tuning with nearest-template reasoning. “A Template Is All You Meme” constructs a knowledge base of 5,220 base templates and 49,531 example instances, encodes them with CLIP, and classifies memes by nearest-template lookup and template-majority labels. Its Template-Label Counter outperforms prior fine-tuned baselines on several datasets, including MultiOff and Memotion 3 tasks, and motivates template-aware splits that prevent train/test leakage across templates (Bates et al., 2023).

Retrieval-augmented and evidence-oriented systems push beyond direct classification. MEMEX frames meme understanding as sentence-level evidence detection over related documents; its MIME model reaches test-set F1 0.812, compared with 0.772 for the strongest baseline MMBT (Sharma et al., 2023). Open-world systems extend this idea to recent memes whose meaning is absent from model parameters: Query-Retrieve-Conclude identifies missing knowledge, retrieves open-web evidence, synthesizes background knowledge statements, and improves evidence recall on a recent Know Your Meme benchmark from 0.46 to 0.78 in one reported setting, while raising downstream overall macro-F1 from 0.65 for vanilla zero-shot detection to 0.71 (Liu et al., 3 Jun 2026).

5. Affect, explanation, and contextual reasoning

Emotion and affect are recurring targets of meme analysis, but the modeling assumptions vary. A text-only DistilBERT classifier applied to meme text predicts six basic emotions—joy, anger, fear, sadness, surprise, and love—and, on the Memotion dataset, yields the distribution Joy 44.65% (3120), Anger 33.98% (2374), Fear 9.00% (629), Sadness 8.97% (627), Surprise 2.05% (143), and Love 1.35% (94). The same work reports a statistically significant association between emotion and motivational content with $p=0.045$, and a user-study agreement rate of 67.23% between CLIP-based similarity grouping and human judgments across 21 meme groups (Konyspay et al., 21 Mar 2025). More recent meme-emotion systems move beyond text-only inference: MemoDetector uses four-step MLLM-based textual enhancement and dual-stage fusion, improving F1 scores by 4.3% on MET-MEME and 3.4% on MOOD (Shi et al., 14 Nov 2025).

Context-sensitive reasoning exposes a harder layer of the problem. MemeReaCon preserves meme image, post text, and top comment, annotating five dimensions including context-meme interplay, meme type, comment stance and affective consistence, post connection, and post intent. On this benchmark, the best reported model, Gemini-2.5-Pro, reaches 83.21% accuracy on context-meme interplay classification and 71.28% on stance-plus-affect classification, but only 60.38 ROUGE-L on post-connection generation and 44.86 ROUGE-L on post-intent generation, indicating a gap between surface relation recognition and full pragmatic synthesis (2505.17433).

Explanation-oriented work treats interpretability as a first-class output. ExHVV augments hero-villain-victim role labeling with natural-language explanations for 4,680 entity instances in 3K memes, and LUMEN improves over the best baseline across 18 generation metrics, reaching BLEU-4 0.313, ROUGE-L 0.530, and CIDEr 1.380 (Sharma et al., 2022). MemeCap similarly shows that even strong vision-language systems still struggle with visual metaphors and perform substantially worse than humans, especially when they literalize vehicles, copy embedded text, or hallucinate unsupported content (2305.13703). Taken together, these results suggest that affect, stance, and explanation are not separable add-ons but core dimensions of meme understanding.

6. Generation, transformation, and applied uses

Generative work typically separates template choice from caption production. memeBot formulates meme generation as

$$p(y,I\mid x)=p(I\mid x;\phi)\cdot p(y\mid x,I;\theta),$$

selecting one of 24 templates and then generating a caption conditioned on both the input sentence and selected template. On Twitter-based evaluation, it reports coherence 2.66/4, relevance 2.65/4, and user likes 0.65 (Sadasivam et al., 2020). A related text-driven system first predicts input emotion and then selects a template, with its best emotion classifier—a fine-tuned BERT model—reaching Accuracy 83.29% and F1 82.37%; human evaluation of its caption generator over 100 test sentences yields 62 Good, 26 Acceptable, and 12 Bad outputs (Liu et al., 2021).

Template-conditioned large-scale generation is represented by Memeify, which assembles 1.1 million meme captions from 128 classes and organizes them into six themes. In qualitative evaluation, average ratings are 3.23 for original memes, 3.10 for Memeify-generated memes, and 2.98 for the baseline; in a generated-versus-original discrimination task, participants misclassified Memeify-generated memes as original 66.67% of the time (Vyalla et al., 2019). Stance-aware systems such as MemeCraft add social cause, stance, and persuasion technique as controls, and apply a hateful-meme filter with threshold $\tau=0.9$; its ChatGPT-based variant substantially outperforms the Dank Learning baseline on authenticity while keeping residual human-judged hatefulness low after filtering (Wang et al., 2024).

More recent work addresses controlled multimodal transformation rather than de novo generation. MER-Bench defines “Meme Reappraisal,” in which a negative meme is transformed into a constructive one while preserving scenario, entities, and structural layout. It introduces the Reappraisal Fidelity Score and evaluates 14 editing or generation systems; the best reported result is Flux9B with RFS $76.78\pm1.16$, while many systems either preserve structure without shifting affect or shift affect while destroying meme layout (Nie et al., 16 Mar 2026). This establishes meme editing as a distinct problem from ordinary image generation or caption generation.

In application terms, the literature repeatedly identifies moderation, retrieval, trend tracking, explanation, and media-literacy support as primary use cases. Similarity clustering has been proposed for near-duplicate detection and policy enforcement, retrieval models such as mtrCLIP improve meme-caption and embedded-text retrieval, and explanation systems are positioned as aids for content moderation and integrity analysis (Konyspay et al., 21 Mar 2025, Deng et al., 23 Jan 2025, Sharma et al., 2023).

7. Limitations, controversies, and research directions

A central controversy concerns what counts as a meme in the first place. A memetics-aware audit of seven leading meme classification datasets found that 50.4% of evaluated samples contained no signs of memetics, and that none of 12 qualitatively examined datasets reported a replicable meme-identification methodology (Hazman et al., 2024). This finding directly challenges any benchmark that treats “image with text” as equivalent to “meme,” and it motivates stricter dataset construction based on verifiable memetic elements and related-but-distinct instances.

Methodologically, recurrent limitations include missing OCR, narrow linguistic coverage, label noise, weak cultural grounding, and incomplete evaluation. In multimodal similarity work, no OCR is applied to the Reddit Memes Dataset and the reported participant count is inconsistent between abstract and results; in template-based classification, standard train/test splits allow template leakage; in MemeCap, state-of-the-art models still fall substantially short of humans on correctness, completeness, and faithfulness (Konyspay et al., 21 Mar 2025, Bates et al., 2023, 2305.13703). These issues suggest that reported gains may reflect shortcut learning, dataset idiosyncrasies, or restricted cultural scope as much as genuine meme understanding.

The main research directions are correspondingly clear. Several works call for robust OCR and layout-aware modeling, stronger multimodal fusion such as cross-attention or gated fusion, multilingual and cross-cultural support, context-aware reasoning over conversations and communities, template-aware splits, and retrieval-augmented or open-world knowledge acquisition for emerging memes (Konyspay et al., 21 Mar 2025, Hazman et al., 2024, Liu et al., 3 Jun 2026, Shahroor et al., 18 Jan 2026). A plausible implication is that progress on memes will depend less on a single universal architecture than on combining memetics-aware curation, explicit multimodal grounding, temporal knowledge access, and evaluation protocols that distinguish template familiarity from actual interpretive competence.