Graffiti: Urban Markings & Art Forms

• Graffiti is a heterogeneous phenomenon of unsanctioned urban markings, heritage inscriptions, and activist public art, reflecting varied legal and cultural contexts.
• Research classifies graffiti into distinct types—simple scribbles, complex writings, and textured canvases—using spatial analyses and image-based methodologies.
• Computational models, from CNNs to agent-based simulations, demonstrate graffiti’s role in territorial dynamics, pattern recognition, and security challenges.

Graffiti denotes a heterogeneous class of marks, inscriptions, and images placed in public or semi-public space. In the cited research, the term spans unsanctioned urban markings and graffiti art, activist public art, historical carved inscriptions, gang territorial markings, and digitally generated or robotically reproduced traces. This multiplicity is not merely terminological: it determines the relevant ontology, data modality, and analytical method. São Paulo street-view studies treat graffiti as a city-scale spatial phenomenon with distinct visual categories (Tokuda et al., 2020); heritage-computing work treats medieval carved letters from St. Sophia Cathedral in Kyiv as a difficult epigraphic recognition problem (Gordienko et al., 2018, Gordienko et al., 2018); territoriality models formalize graffiti as a mediating field that stores local memory and induces segregation (Barbaro et al., 2012, Alsenafi et al., 2018, Alsenafi et al., 2021); and recent AR, diffusion, and robotics systems treat graffiti as embodied public art, a trajectory style, or a high-contrast generative medium (Jiang, 2024, Berio et al., 2017, Chen et al., 2021, Berio et al., 3 Jul 2025, Banerjee et al., 28 Aug 2025).

1. Conceptual scope and typological distinctions

A recurrent finding in the literature is that graffiti is not a single homogeneous category. One strand distinguishes sanctioned public art, such as commissioned murals, from unsanctioned work, such as graffiti art made without permission; this distinction matters because legality, anonymity, and site-specificity shape both interpretation and technological design (Jiang, 2024). Another strand divides urban graffiti by visual form into simple scribbles (Type A), complex scribbles / painted writings (Type B), and canvases or textures (Type C), arguing that differences in form, effort, and execution may correspond to different spatial behaviors (Tokuda et al., 2020). A third strand uses “graffiti” in the historical-epigraphic sense, referring to Glagolitic and Cyrillic letters carved into cathedral walls between the 11th and 18th centuries (Gordienko et al., 2018).

Research framing	Core object	Representative source
Urban public marking	Type A, B, C graffiti in street-view imagery	(Tokuda et al., 2020)
Activist public art	Site-specific public art, including graffiti art without permission	(Jiang, 2024)
Historical inscription	Carved Glagolitic and Cyrillic letters	(Gordienko et al., 2018)
Territorial marker	Graffiti fields mediating gang interactions	(Barbaro et al., 2012)
Embodied trajectory	Graffiti-like strokes and motion style	(Berio et al., 2017)

This classification suggests that any general theory of graffiti must distinguish at least among inscription, image, social act, spatial signal, and motor trace. A plausible implication is that disagreements over whether graffiti is vandalism, art, or data often arise because different studies hold different ontologies fixed.

2. Historical graffiti, epigraphy, and recognition

In heritage and pattern-recognition research, graffiti refers to carved historical handwriting rather than aerosol painting. The CGCL dataset—carved Glagolitic and Cyrillic letters from the stone walls of St. Sophia Cathedral in Kyiv—contains more than 4,000 images spanning 34 letter classes and is presented as an open source resource for machine learning. The cited study emphasizes that these inscriptions, dated from the 11th to 18th centuries, are visually degraded, irregular, and often incomplete; it further notes that graffiti dated to about 1018–1022 are cited as confirming the cathedral’s foundation in 1011. Exploratory analysis with t-SNE indicates that CGCL glyphs are less separable than cleaner font-derived glyphs in notMNIST. As a baseline, multinomial logistic regression achieves per-class AUC values not lower than 0.60 on CGCL, with averaged AUC 0.82; a 2D CNN with 5 convolutional/max-pooling layers, 205,217 trainable parameters, ReLU, binary cross-entropy, and RMSProp at learning rate $10^{-4}$ reaches, under high lossy data augmentation, test accuracy 0.94, loss 0.21, and AUC 0.99 on the CGCL A/H subset (Gordienko et al., 2018).

A subsequent study compares a capsule deep neural network with a previously used CNN on the same historical domain. On A/H subsets with 180 training images, 40 validation images, and 70 test images, the capsule model is reported to train 5–6 times faster, avoid overtraining, and produce better validation accuracy, validation loss, ROC-AUC, and confusion matrices. Reported capsule-network AUC ranges are 0.88–0.93 without augmentation, 0.91–0.95 with lossless augmentation, and 0.91–0.93 with lossy augmentation, compared with CNN values of 0.50, 0.51, and about 0.90, respectively. The architecture includes convolutional layers, primary capsules, digit capsules, and a decoder branch, with more than 80 million trainable parameters, ReLU and sigmoid activations, binary cross-entropy, Adam, and learning rate $10^{-4}$ (Gordienko et al., 2018).

These results establish historical graffiti as a technically demanding recognition domain in which erosion, incomplete letter forms, and noncanonical geometry are central, not incidental, features.

3. Urban spatial distribution and network structure

A city-scale study of São Paulo operationalizes graffiti through street-view imagery and city topology. Using a street network extracted from OpenStreetMap and 16,000 randomly sampled Google Street View images, the study finds that 3,154 images contain at least one graffiti type. Type counts are 2,086 for Type A, 476 for Type B, and 592 for Type C, with multiple types allowed in the same image (Tokuda et al., 2020).

Spatial concentration is estimated with kernel density using a Gaussian kernel. To avoid overinterpreting sparse regions, locations below one occurrence per 500 m are excluded. Relative to the mean distribution, Type A is less concentrated in the northeast of the city, Type B is more concentrated there, and Type C exhibits two strong concentration areas in the eastern region. Distributional distinctness is quantified with the Kullback–Leibler divergence

$$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$

with reported values 0.0137 for Type A, 0.0141 for Type B, and 0.009 for Type C. The study interprets these values as indicating that Type A and Type B differ more from the mean than Type C, with Type A the most distinct (Tokuda et al., 2020).

The same work embeds graffiti in the street network itself. Infomap partitions São Paulo into six communities based on random-walk compression, after which type prevalence is examined per community. Total graffiti counts vary regionally, but the relative proportions of Types A, B, and C do not differ significantly across communities. A second network descriptor, outward accessibility,

$$A^{out}_{h}(i)=\exp\left[-\sum_{j=1}^{n} p^h_{i,j}\log(p^h_{i,j})\right],$$

is computed with $h=20$ self-avoiding random-walk steps. Accessibility is lower near the city border and higher in central regions, and Pearson correlations between accessibility and graffiti occurrence are 0.34 for Type A, 0.35 for Type B, and 0.30 for Type C, characterized as small positive correlations (Tokuda et al., 2020).

The significance of this line of work lies in its refusal to treat “graffiti” as a unitary nuisance variable. Form, network embedding, and sampling geometry all affect the observed distribution.

4. Graffiti as territorial field and mathematical mechanism

In mathematical sociology and statistical mechanics, graffiti is modeled not as image content but as an interaction field. A foundational lattice model introduces agent variables $\eta_i \in \{-1,0,+1\}$ and continuous graffiti variables $g_i\in\mathbb{R}$, with positive $g_i$ representing excess red graffiti and negative $g_i$ excess blue graffiti. The GI-Hamiltonian

$$- \mathscr{H}(\mathbf{s}) = J \sum_{\langle i,j\rangle} \eta_i g_j + K \sum_i \eta_i g_i + \alpha \sum_i \eta_i^2 - \lambda \sum_i g_i^2$$

contains no direct gang-to-gang term. The cited analysis concludes that clustering and territory formation can arise solely from gang-to-graffiti couplings. In the mean-field rendition, the transition is continuous when $10^{-4}$0 and discontinuous when $10^{-4}$1, with tricritical point

$10^{-4}$2

This is presented as evidence that direct gang-to-gang interactions are not strictly necessary for territory formation (Barbaro et al., 2012).

A related two-gang agent-based model makes the mechanism dynamic. On a 2D lattice, agents move by biased random walk and avoid rival graffiti with transition weights exponentially suppressed by the enemy field. Graffiti is deposited at rate $10^{-4}$3 and decays at rate $10^{-4}$4, producing the continuum system

$10^{-4}$5

$10^{-4}$6

Linear stability analysis yields a precise instability threshold for the well-mixed state: $10^{-4}$7 This shows that segregation is promoted by high graffiti production, slow decay, and higher agent density (Alsenafi et al., 2018).

The $10^{-4}$8-group generalization replaces two graffiti fields with $10^{-4}$9 markings and produces a system of $$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$0 convection-diffusion equations. The total opposing graffiti seen by group $$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$1 is

$$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$2

and the corresponding continuum equation is

$$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$3

For three groups, the paper reports a discrete critical value around

$$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$4

It also studies Timidity and Threat Level variants, which relocate the parameter $$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$5 from the moving gang to the marking gang and thereby generate distinct segregation geometries (Alsenafi et al., 2021).

These models formalize graffiti as dynamic memory: marks persist, decay, bias future motion, and amplify small asymmetries into territorial structure.

5. Public art, activism, accessibility, and embodied AR

Another body of research treats graffiti as site-specific public art, often created without permission and frequently bound to political expression. A position paper on AR-enabled accessibility argues that public art can be both artifact and activism, and uses graffiti as a key example of unsanctioned, often anonymous, public expression. The proposed ARtivism process artifact includes an early audiovisual prototype, a labeled map of mural locations, and a soundscape/video prototype. It builds on a prior crowdsourced description project with 113 descriptions for 14 works of public art submitted by 25 participants across the U.S. and Canada. The prototype surfaces descriptions through AR with a green icon indicating description availability, a text box naming the author or narrator, and plus/minus zoom controls; the associated site map uses accessible thumbnails with alt text giving the artwork’s name, a brief high-level description, and approximate location. The paper emphasizes tensions between access and artist privacy, particularly for unsanctioned graffiti whose documentation may expose location and authorship (Jiang, 2024).

A later AR system, GestoBrush, shifts attention from accessibility for audiences to embodied creation for graffiti artists. Built as a Unity-based iOS app using Apple ARKit, RealityKit, a Golang server, and HTTP, it turns the smartphone into a virtual spray can whose edge functions as a nozzle. Users long-press a central button to draw, first on a scanned wall and then into 3D space beyond the wall. The system provides two virtual tools, a graffiti spray and a drip mop, and stores/retrieves models using a backend while rendering user-created graffiti with a triangular mesh algorithm parameterized by position, size, and color. A co-design workshop with 5 graffiti artists informed the design, and an evaluation with 6 graffiti artists in Tianjin reported enhanced intuitiveness, immersion, expressiveness, and the ability to bypass real-world constraints. The same study also reports a learning curve for 3D control, fatigue, scale-placement difficulty, and unstable tracking in bright or low-feature environments (Chen et al., 6 Sep 2025).

Together these systems show that digital graffiti research is not confined to image synthesis. It also concerns description, bodily practice, spatial context, and the ethics of cataloging unsanctioned work.

6. Computational stylization, generation, and robotic execution

A trajectory-based strand models graffiti as motor behavior. One framework learns stylization patterns from a small number of example drawings or writings using a Recurrent Mixture Density Network (RMDN) combined with the Sigma Lognormal model of movement. Traces are decomposed into a structural layer of virtual targets and a dynamic layer of stroke parameters, and the network predicts distributions over $$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$6. The model uses 2 recurrent layers, 400 units per layer, 20 Gaussians, dropout keep probability 90%, Adam, and truncated BPTT. The method supports augmentation factors such as $$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$7 for one-example learning and is explicitly demonstrated on calligraphy- and graffiti-like trajectories, including asemic graffiti generation (Berio et al., 2017).

A portrait-generation strand treats graffiti as a high-contrast visual style that threatens identity preservation. CraftGraffiti uses a LoRA-fine-tuned diffusion transformer on the 17K-Graffiti dataset, then applies a face-consistent self-attention mechanism with identity embeddings and CLIP-guided prompt extension for pose control. The paper justifies a “style-first, identity-after” ordering and reports for the full system FFC = 0.7713, Aes = 5.2271, HPS = 0.3536, and Inference time = 10.1 s. A human study with 47 anonymous users evaluating 60 outputs ranks the method best in aesthetics, style blending, and recognizability (Banerjee et al., 28 Aug 2025).

A motion-aware robot-drawing line brings graffiti into physical execution. One method combines the sigma-lognormal model with DiffVG so that image-space losses can optimize human-like movements directly; the objective combines image fidelity with a minimum-time smoothing term, and the implementation uses PyTorch, Adam, and 300 steps. The paper demonstrates synthetic graffiti generation and robotic reproduction, arguing that graffiti is better understood as “urban calligraphy” than as static vector geometry (Berio et al., 3 Jul 2025). A complementary system, GTGraffiti, captures real human painting motions with an OptiTrack system at 120 Hz, uses a purpose-built planar 4-cable cable-driven parallel robot, and converts artist paths into executable trajectories with offline iLQR and an online time-varying feedback law. The robot reproduces motions up to 2 m/s and 20 m/s$$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$8 with 9.3 mm RMSE; hardware capability reaches 7.6 m/s and 94 m/s$$D_\text{KL}(p(x)\parallel q(x))=\int_{-\infty}^{\infty} p(x)\log\left(\frac{p(x)}{q(x)}\right)\,dx,$$9 per winch, while the full frame measures 3.05 m × 2.44 m × 0.61 m (Chen et al., 2021).

This computational literature shifts the analytic focus from the finished mark to the generative procedure—stylistic, kinematic, or control-theoretic—through which graffiti is produced.

7. Graffiti as corruption, control signal, and security threat

Recent machine-learning work also treats graffiti as an occlusion process, a control modality, or an adversarial trigger. In text-image restoration, graffiti-like scribbles are modeled as corrosion or irregular overlay that destroys the global structure of text. To study this, one paper introduces TII-ST with 86,476 images—80,000 synthesized scene text images plus 6,476 real images—and TII-HT with 40,078 IAM handwriting images, using Convex Hull (CH), Irregular Region (IR), and Quick Draw (QD) corruptions with ratios from 5%–60%. Its Global Structure-guided Diffusion Model (GSDM) first predicts a segmentation prior and then reconstructs the full image. On TII-ST it reports CRNN 67.48%, ASTER 74.67%, MORAN 73.04%, PSNR 33.28 dB, and SSIM 0.9596; on TII-HT it reports DAN 69.43%, TrOCR-Base 56.00%, TrOCR-Large 66.81%, PSNR 32.13 dB, and SSIM 0.9718 (Zhu et al., 2024).

A different paper treats ancient graffiti on rock as a zero-shot associative generation task. Its self-supervised controllable generation (SCG) framework learns modular visual factors from MS-COCO only, then uses a learned module such as HC3 to condition a diffusion model without graffiti-specific supervision. In qualitative comparisons against ControlNet conditioned by Canny edges, SCG is reported to suppress rock-surface noise more effectively and produce more natural outputs; subjective evaluation on ancient graffiti uses 37 participants, with similar winning rate on fidelity but significantly higher winning rate on aesthetics (Chen et al., 2024).

In security research, graffiti becomes an attack primitive. A multimodal backdoor method for autonomous-driving VLMs generates graffiti-based visual patterns via stable diffusion inpainting and pairs them with a cross-lingual text trigger. On DriveVLM-Base, the method reports average ASR 86.67% and average FPR 0.19%; on DriveVLM-Large, average ASR 90.00% and average FPR 0.00%. At poisoning ratios $$A^{out}_{h}(i)=\exp\left[-\sum_{j=1}^{n} p^h_{i,j}\log(p^h_{i,j})\right],$$0, the large model reaches 75.00 / 95.00 / 100.00 ASR with 0.00 / 0.00 / 0.00 FPR. The trigger is designed for non-salient planar regions such as walls and barriers, explicitly excluding vehicles and pedestrians, and is motivated by the fact that graffiti appears natural in urban scenes while remaining semantically non-critical to driving (Wang et al., 6 Apr 2026).

These uses complicate any simple account of graffiti as either nuisance or art. In current technical literature it can also be a corruption prior, a latent structural cue, or a covert signal embedded in safety-critical systems.