CraftGraffiti: Exploring Human Identity with Custom Graffiti Art via Facial-Preserving Diffusion Models

TL;DR

CraftGraffiti is a novel AI framework that generates graffiti art from text prompts, preserving facial identity and allowing pose customization without keypoints, validated through quantitative and real-world tests.

Contribution

It introduces a face-preserving diffusion-based graffiti generation method with pose control, advancing identity fidelity in stylized AI art.

Findings

Achieves high facial feature consistency in graffiti images.

Outperforms existing methods in aesthetic and human preference scores.

Demonstrates real-world creative impact through festival deployment.

Abstract

Preserving facial identity under extreme stylistic transformation remains a major challenge in generative art. In graffiti, a high-contrast, abstract medium, subtle distortions to the eyes, nose, or mouth can erase the subject's recognizability, undermining both personal and cultural authenticity. We present CraftGraffiti, an end-to-end text-guided graffiti generation framework designed with facial feature preservation as a primary objective. Given an input image and a style and pose descriptive prompt, CraftGraffiti first applies graffiti style transfer via LoRA-fine-tuned pretrained diffusion transformer, then enforces identity fidelity through a face-consistent self-attention mechanism that augments attention layers with explicit identity embeddings. Pose customization is achieved without keypoints, using CLIP-guided prompt extension to enable dynamic re-posing while retaining facial coherence. We formally justify and empirically validate the "style-first, identity-after" paradigm, showing it reduces attribute drift compared to the reverse order. Quantitative results demonstrate competitive facial feature consistency and state-of-the-art aesthetic and human preference scores, while qualitative analyses and a live deployment at the Cruilla Festival highlight the system's real-world creative impact. CraftGraffiti advances the goal of identity-respectful AI-assisted artistry, offering a principled approach for blending stylistic freedom with recognizability in creative AI applications.

Tables

Table 1. Table 1: Quantitative Evaluation of CraftGraffiti

Method/Metric	FFC $\uparrow$	Aes $\uparrow$	HPS $\uparrow$	Inf. Time (sec) $\downarrow$
FLUX + IP-Adapter (Ye et al., 2023)	0.8324	3.2414	0.3012	8.2
FLUX.1 Kontext (Labs et al., 2025)	0.6741	2.1749	0.2911	6.7
InstructPix2Pix (Brooks et al., 2023)	0.7112	2.7272	0.1918	3.1
GPT4o (Hurst et al., 2024)	0.8761	4.5193	0.3412	13.4
Grok 3 (Grok, 3)	0.8513	4.1652	0.3102	11.2
Ours (Baseline)	0.7618	3.6913	0.2911	2.9
Ours (+ Style Fusion)	0.6814	4.7195	0.3001	5.3
Ours (+ Face Consistent self attention)	0.7713	5.1376	0.3176	8.7
CraftGraffiti	0.7713	5.2271	0.3536	10.1

Equations

$$q({\mathbf{x}}_t\mid {\mathbf{x}}_{t-1})=\mathcal{N}({\mathbf{x}}_t;\sqrt{1-{\beta }_t}{\mathbf{x}}_{t-1},{\beta }_t\mathbf{I}),$$

$$p_{\theta }({\mathbf{x}}_{t-1}\mid {\mathbf{x}}_t,\mathbf{c})=\mathcal{N}({\mathbf{x}}_{t-1};{\mathbf{\mu }}_{\theta }({\mathbf{x}}_t,t,\mathbf{c}),{\mathbf{\Sigma }}_{\theta }({\mathbf{x}}_t,t,\mathbf{c})),$$

$${\mathcal{L}}_{\text{style}}(\mathbf{x},\mathbf{s})=\sum _l{\Vert G^{(l)}({\varphi }^{(l)}(\mathbf{x}))-G^{(l)}({\varphi }^{(l)}(\mathbf{s}))\Vert }_F^2,$$

$${\mathcal{L}}_{\text{content}}(\mathbf{x},\mathbf{c})={\Vert {\varphi }^{(l)}(\mathbf{x})-{\varphi }^{(l)}(\mathbf{c})\Vert }_2^2,$$

$${\mathcal{L}}_{\text{total}}={\lambda }_c{\mathcal{L}}_{\text{content}}+{\lambda }_s{\mathcal{L}}_{\text{style}},$$

$$\tilde{W}=W+\alpha \cdot AB,$$

$$\mathbf{u},\mathbf{v}\in {\mathbb{R}}^n$$

$$\text{cosine\_similarity}(\mathbf{u},\mathbf{v})=\frac{\mathbf{u}\cdot \mathbf{v}}{\Vert \mathbf{u}{\Vert }_2\Vert \mathbf{v}{\Vert }_2}$$

Full text

1 Introduction

Human identity preservation in the GenAI era: bias, aesthetics, and open experimentation: Generative systems such as GANs and diffusion models replicate and amplify cultural biases from their training data, entering into the risk of distorting human identity across multiple dimensions (Ghosh et al., 2024). Since usually trained on not particularly curated datasets, it is known that current generative systems usually tend to disproportionately produce light-skinned, youthful, conventionally attractive faces, marginalizing other demographics, and concentrating on a narrow band of a specific concept of aesthetics (Muñoz et al., 2023). On the other hand, both children and older adults are usually either underrepresented or not properly identified, even for age-progression models (Muñoz et al., 2023; Li et al., 2018), presenting results that tend to transform the image of a young person into an adult face, or conversely, elderly people as unrealistically younger figures. Errors in the gender identification are usually more prone to affect women than men, and reinforce gender norms and heteronormativity, as seen in occupational and presentational stereotypes (Sun et al., 2023; Zhou, 2024). Rooted in imbalanced datasets, these distortions risk homogenizing outputs unless addressed through curated data, fairness-aware training, and interdisciplinary critique (Ferrara, 2023).

For this reason, and in light of the relevant risks identified, it is essential to openly address the tensions to which our human identity is exposed in the new digital era; an era of interconnected individuals endowed with generative capacities. Such a topic is profound and complex from the philosophical, ethical, political, and technical dimensions (Binns, 2017; Al-kfairy et al., 2024). We state that in order to address the discussion and public debate of these topics, we can make use of two main instruments: 1) An artistic installation providing a representation of the customized human pose, technically sound and preserving facial attributes as much as possible, and 2) an ecological environment that allows for the natural interaction with people, providing the explicit context of a living lab or open experimentation space.

CraftGraffiti** - Integrating Cultural Expression and Facial Consistency in Generative Street Art:** For the first aim, we need a cultural bind that serves as an entry point for people to interact with this new era of GenAI, which has to be inclusive and accessible for all; from this perspective, graffiti appears as an excellent candidate. Graffiti is a culturally significant art form, merging creative expression with the diversity of human identity (Das, 2023; Forster et al., 2012) and often serving as a voice for marginalized communities (Bates, 2014). While recent advances in GANs (Xu et al., 2021; Yang et al., 2019) and diffusion models (Zhang et al., 2023a; Wang et al., 2023b) offer automated generation, existing methods struggle to preserve facial attributes across varying styles and poses. Text-based image editing (Kawar et al., 2023; Huang et al., 2025) and style transfer models (Huang et al., 2024; Wang et al., 2023b; Guo and Lin, 2024) fail to simultaneously adjust poses, while pose editing approaches (Cheong et al., 2023; Okuyama et al., 2024) lack unified style transfer. Unified models (Pascual et al., 2024; Mokady et al., 2023; Dong et al., 2023) attempt both but often lose facial details due to self-attention limitations.

We introduce CraftGraffiti, an end-to-end generative framework that produces custom graffiti-style portraits of human characters while explicitly preserving the subject’s facial identity (see Fig.˜1). This system is text-guided, allowing a user to specify high-level context (e.g., prompting a portrait of the person as a DJ, guitarist, or singer) and to customize the character’s pose, all within a graffiti art style. To accomplish this, CraftGraffiti integrates several components: a CLIP-based text encoder for semantic guidance, a graffiti-style fusion module implemented via a pre-trained diffusion transformer 111https://civitai.com/models/1058970/graffiti-style-fluxFaces fine-tuned on graffiti aesthetics, a LoRA-based adapter (Hu et al., 2022) that efficiently injects pose information without requiring full model retraining, and a novel face-consistent self-attention mechanism that ensures key facial features of the input are maintained throughout the diffusion process. The final output is rendered by a variational autoencoder (VAE) decoder, yielding a high-quality graffiti portrait that retains the individual’s identity. This architecture enables simultaneous pose adjustment and style transfer while keeping the person’s face recognizable. Our contributions include: (1) a unified framework for pose-guided, graffiti-themed portrait generation that preserves individual identity; (2) a novel self-attention module specifically designed for facial consistency in diffusion-based generation; and (3) state-of-the-art performance in both qualitative visual fidelity and quantitative identity-preservation metrics compared to existing methods.

A public artistic installation in an ecological environment as a living lab: For our second aim, we propose to explore ecological environments to enhance user-experience experimentation by enabling real-world engagement in authentic contexts (Kieffer, 2017). Living labs provide user-centred, open innovation ecosystems with real-world experimentation, co-creation, and multi-stakeholder collaboration (8; 29). Outdoor labs with pervasive sensing and crowdsensing further increase scalability and context-awareness (Pournaras et al., 2021), fostering inclusivity, diversity, and richer insights than artificial environments. From this perspective, music festivals appear as excellent ecological experimentation spaces for this type of trans- and multi-disciplinary investigation. In our case, we set our installation during the four days of the Festival Cruïlla, taking place 9-12 July 2025 (Please refer to the appendix for more details of the installation setup, outcomes, and human feedback).

2 Preliminaries

Diffusion Models: Text-to-image diffusion models, such as Stable Diffusion Rombach et al. (2022); Podell et al. (2023); Everaert et al. (2023), generate images by learning a denoising process that maps latent noise to realistic images ($\mathbf{I}$), conditioned on textual input. Let $\mathbf{x}_{0}\in\mathbb{R}^{H\times W\times C}$ denote a clean image and $\mathbf{x}_{t}$ its noisy version at timestep $t\in\{1,\dots,T\}$. The forward diffusion process ($q(\cdot)$) adds Gaussian noise ($\mathcal{N}(\cdot)$) in a Markov chain:

[TABLE]

with a predefined variance schedule $\{\beta_{t}\}$. The model learns the conditional reverse process ($p_{\theta}$):

[TABLE]

where $\mathbf{c}$ is a conditioning vector, often a CLIP-encoded text prompt. Training minimizes the error between predicted and true noise or the original image $\mathbf{x}_{0}$, depending on parameterization (mean: $\boldsymbol{\mu}_{\theta}$, and covariance $\boldsymbol{\Sigma}_{\theta}$). Latent diffusion models (LDMs) (Wu and De la Torre, 2023; Nam et al., 2024; Zhang et al., 2023b) perform this process in a compressed latent space $\mathcal{Z}$, using an encoder $\mathcal{E}:\mathbb{R}^{H\times W\times C}\rightarrow\mathcal{Z}$ and decoder $\mathcal{D}:\mathcal{Z}\rightarrow\mathbb{R}^{H\times W\times C}$, improving computational efficiency while preserving generation quality.

Style Transfer: Style transfer (Wang et al., 2023a; Hamazaspyan and Navasardyan, 2023; Everaert et al., 2023) aims to modify the appearance of a generated image $\mathbf{x}$ to match the style of a reference image $\mathbf{s}$, while maintaining the content from a source image $\mathbf{c}$. In feature space $\phi(\cdot)$, the style and content losses are defined as:

[TABLE]

[TABLE]

where $G^{(l)}$ denotes the Gram matrix Li et al. (2017) of features at layer $l$, capturing second-order statistics. The total loss combines both objectives:

[TABLE]

with tunable weights $\lambda_{c}$ and $\lambda_{s}$. Modern approaches integrate style conditioning into generative models via cross-attention modulation or adapter modules. We fine-tune the transformer of the FLUX.1 dev 222https://huggingface.co/black-forest-labs/FLUX.1-dev with low-rank adaptation via the $\mathcal{L}_{\text{total}}$ in order to learn the graffiti style and use it as a pre-trained model during inference to pose the style over the input image (see Fig.˜2).

Pose Customization with LoRA: Generation of characterization conditions on a specified human pose $\mathbf{p}$, often represented as a 2D keypoint skeleton or heatmap (Wang et al., 2024b; Ju et al., 2023; Khandelwal, 2024). This signal can be embedded and injected into a diffusion model via attention layers or early-stage concatenation. LoRA (Hu et al., 2022) offers an efficient means of adapting large diffusion models to new tasks (e.g., pose-guided generation) by injecting low-rank updates into pretrained weight matrices. For a weight matrix $W\in\mathbb{R}^{d\times k}$, LoRA introduces trainable low-rank matrices $A\in\mathbb{R}^{d\times r}$, $B\in\mathbb{R}^{r\times k}$, modifying the weights as:

[TABLE]

where $\alpha$ is a scaling factor and $r\ll\min(d,k)$. By freezing base weights and only training $A$ and $B$, LoRA allows efficient fine-tuning with significantly fewer parameters. LoRA can be applied to attention matrices $W_{q},W_{k},W_{v}$, enabling fast pose-driven customization without full model retraining. In CraftGraffiti instead of using keypoints or heatmap, we extend the input prompt via T5 (Narang et al., 2021) and get the cross-attention map as a guided signal ($B$) for the pose (see Fig.˜2). We maintain the identity of the character through the face-consistent self-attention ($A$) and update the self and cross attention of the diffusion transformer via fine-tuning with LoRA for multi-scale latent feature generation. This multi-scale latent feature is further denoised via VAE in order to get the final edited image in graffiti style with an accurate musical pose described in the text prompt.

3 CraftGraffiti

3.1 Why should we perform graffiti style transfer before facial attribute preservation

Performing the style transform before enforcing attribute constraints can be seen via a simple model. Let $I$ be the original face image and let $S(I)$ be its graffiti version under a LoRA-fine-tuned diffusion model. If we have a facial-attribute extractor $F_{a}(I)\in\mathbb{R}^{k}$ (encoding eyes, nose, chin, expression, etc.) and an attribute-loss $\mathcal{L}_{attr}(X)=||F_{a}(X)-F_{a}(I)||^{2}$, we can choose a projection $P(\cdot)$ so that for any image $X$, $F_{a}(P(X))=F_{a}(I)$ (i.e. $P$ restores $I$’s attributes). Then by reconstruction $L_{\rm attr}(P(S(I)))=0$, whereas $P(I)=I$ (since $I$ already has its own attributes), so $\mathcal{L}(S(P(I)))=\mathcal{L}(S(I))=||F_{a}(S(I))-F_{a}(I)||^{2}>0$ whenever $S$ perturbs the attributes. In other words, $\mathcal{L}_{attr}(P\circ S(I))\leq\mathcal{L}_{attr}(S\circ P(I))$. This shows formally that applying the style $S$ first and then the attribute correction $P$ yields lower attribute discrepancy. This matches intuition from prior work: style‐transfer methods aim to alter appearance while preserving its underlying structure (Rezaei et al., 2025); however, if the style influence is too strong, it undermines the structural integrity of the content (Wang et al., 2024a). Moreover, LoRA fine-tuned diffusion models often use the early timesteps to reconstruct coarse content and later steps to add stylistic detail (Ouyang et al., 2025). Thus, we preserve identity better by first adding the graffiti style and finally preserving the facial features, rather than applying constraints before styling. The overall methodology of CraftGraffiti has been demonstrated in Fig.˜2.

Theorem. Let $I$ be the input face with attributes $a=F_{a}(I)$. Let $S$ be a diffusion‐style operator and $P$ an operator satisfying $F_{a}(P(X))=a$ for all $X$ (so $P$ restores $I$’s attributes). Then the image $X^{\prime}=P(S(I))$ obeys $F_{a}(X^{\prime})=a$ (it exactly preserves the original attributes), whereas $X^{\prime\prime}=S(P(I))$ satisfies $F_{a}(X^{\prime\prime})=F_{a}(S(I))$ and in general $F_{a}(S(I))\neq a$ unless $S$ itself preserved $F_{a}$. Hence $(P\circ S)(I)$ maintains the facial attributes exactly, whereas $(S\circ P)(I)$ need not.

Proof. By assumption $F_{a}(P(X))=a=F_{a}(I)$ for any input $X$. In particular, $F_{a}(P(S(I)))=a$, so $P(S(I))$ has the original attribute vector. On the other hand, since $P(I)=I$ (the identity image already has attributes $a$), we have $S(P(I))=S(I)$ and hence $F_{a}(S(P(I)))=F_{a}(S(I))$. Unless $S$ is attribute-preserving by itself, $F_{a}(S(I))\neq F_{a}(I)$ in general. Therefore, $P(S(I))$ preserves $a$ exactly, while $S(P(I))$ typically does not, completing the proof.

3.2 How the extra dimension in the self-attention helps in facial attribute preservation

In Diffusion models (Zhou et al., 2024; Wang et al., 2024c), each self-attention layer computes attention only among features of the current latent (no cross-latent computation). Concretely, the latent feature map of size $H\times W$ is reshaped into a sequence of tokens, and for each token, a query ($Q$), key ($K$), and value ($V$) vector are computed from the same latent features. The attention weights are then obtained by taking dot products of $Q$ and $K$ across all positions (scaled by affinity distance $\sqrt{d}$) and passing through softmax, and the result reweights the $V$ vectors (see Fig.˜3). This means every pixel in the latent try to see every other pixel in the same latent when computing self-attention.

To ensure that the two generated characters depict the same identity, we introduce an explicit identity embedding into the attention computation. Conceptually, this means adding a special identity vector (often derived from a reference face) as an extra token or feature dimension in the self-attention layer. In practice, one can concatenate an identity embedding to every spatial token (equivalently, add an extra “CLS”-style token) or append an identity channel to the latent features. The result is that queries and keys now carry identity information as well as image content. In effect, each pixel’s query includes a fixed identity code, and each key (or value) can be augmented by the same code, biasing attention towards features matching that identity.

In an identity-enhanced self-attention, the attention matrix is computed using both the spatial features and the identity embedding. For example, one might form $Q^{\prime}=W_{Q}[x;id]$, $K^{\prime}=W_{K}[x;id]$ where x is a spatial feature and id is the identity vector. Then, attention weights come from $Q^{\prime}K^{\prime T}$. This extra dimension (the identity channel or token) causes the softmax to favor aligning features that correspond to the same identity. Architecturally, the U-Net is unchanged except at attention layers: we simply extend the input feature dimension by the identity embedding. The key effect is that the same identity code is shared across the entire image and across images in a batch. As a result, when generating multiple images of the same person, the self-attention layers see the same identity embedding each time, forcing consistent facial features.

4 Experimental Results

Implementation Details: We implement CraftGraffiti in PyTorch Paszke et al. (2019), utilizing a pre-trained .1 dev 333https://huggingface.co/spaces/black-forest-labs/FLUX.1-dev (12 billion parameter) transformer specifically designed for graffiti style transfer. The guidance scale is set to 7.5. For preserving facial attributes while fine-tuning the transformer with low rank adaptation, the subject guidance factor ($\lambda$) is set to 0.95, and the style intensity factor is set to 0.7. We set the rank ($\sigma$) to 64 and the regularization factor ($\alpha$) to 128 while fine-tuning the transformer for pose customization. We run the denoising process with 100 steps by default. We only perform latent composition in the first quarter of the denoising process (the first 25 steps). All experiments ran on a single NVIDIA A6000 48GB GPU.

Dataset and Evaluation Metric: For graffiti style generation, we have used the images of the 17K-Graffiti dataset 444https://github.com/visual-ds/17K-Graffiti to fine-tune our style fusion diffusion with LoRA and use it as a pre-trained model during inference. Also, to further validate CraftGraffiti, we have gathered images of people from our laboratory, the Computer Vision Center, Barcelona 555https://www.cvc.uab.es/, with their consent, and use them in the paper for validation and pose customization. Furthermore, to perform a fair qualitative analysis, we test CraftGraffiti and other state-of-the-art approaches with facial feature consistency (FFC) (implementation in the Appendix), aesthetic score (Aes) (Murray et al., 2012) calculated using LAION aesthetic classifier, and Human Preference Score (HPS) (Wu et al., 2023). Also, we have compared them against the inference time to understand their real-time use case scenario.

Qualitative Analysis: We compare CraftGraffiti against several FLUX-based approaches, FLUX + IP-Adapter (Ye et al., 2023), and FLUX.1 Kontext (Labs et al., 2025), however, both approaches fail to generate the graffiti style. Although FLUX + IP-Adapter (Ye et al., 2023) tried to maintain facial consistency, it was unable to change style. On the other hand, FLUX.1 Kontext (Labs et al., 2025) tries to change the background into a more cartoonist style, and the faces are not consistent. This shows how hard to unify the style transfer and consistency tasks and justify our decision to perform style transfer before facial consistency (see Fig.˜4).

We also compare CraftGraffiti with InstructPix2Pix (Brooks et al., 2023); however, it neither adds objects nor generates high-quality images. Similarly, VLMs (Grok 3 (Grok, 3) and GPT4o (Hurst et al., 2024)) maintain consistency, unable to blend style due to the complexity of graffiti style transfer. This empirically shows the complexity of the problem (facial consistency in graffiti style) we are tackling and demonstrates the correctness of the component choice of CraftGraffitii for accurate style blending by preserving the facial features (For more qualitative examples, please refer to the Fig.˜10 of the appendix.)

Quantitative Analysis: Table˜1 presents a quantitative comparison against the state-of-the-art image editing techniques. CraftGraffiti stands in the fourth position after GPT4 (Hurst et al., 2024), GPT4o (Hurst et al., 2024), and (Ye et al., 2023) respectively, in terms of facial feature consistency (FFC), as it is very difficult to compare the preservation of facial features, as most of the feature extraction models (InceptionResNetV1) consider global features rather than the local ones. However, during style transfer, it is hard to preserve the global features of the input images. As other image editing (Grok, 3; Hurst et al., 2024; Ye et al., 2023) models are unable to blend the graffiti style, preserve the global features, and perform better in this metric.

On the other hand, Aesthetic Score (Murray et al., 2012; Banerjee et al., 2025) and Human Preference Score (Wu et al., 2023) measure how perfectly the models blend the graffiti style with the input images. CraftGraffiti outperforms all the state-of-the-art image editing techniques in both metrics, as they are unable to blend the graffiti style properly over the input images. Last but not least, CraftGraffiti has very little inference time compared to the VLMs (Grok, 3; Hurst et al., 2024), which makes it suitable for real-time applications.

Ablation Studies: In order to understand the importance of style blending before the pose customization and the face-consistency self-attention, we have performed a brief ablation study depicted in Fig.˜5. It has been observed that with the FLUX.1 dev (12B) baseline, we neither achieve consistency nor achieve the style transfer. On the other hand, we have applied the face consistency self-attention before applying the style transfer. It has been observed that it can preserve the facial attribute during pose customization with low-rank adaptation (2nd row of Fig.˜5). Similarly, if we perform style transfer after the pose customization, neither the facial attributes are preserved, even with the presence of face-consistent self-attention, nor does the style reflect graffiti. As depicted in the 3rd row of Fig.˜5, the images look like a cartoonish drawing on a wall rather than a graffiti poster. So the correct paradigm is to perform the style blending first and then pose customization with face consistent self-attention (4th row of Fig.˜5).

Similarly, we perform an ablation study on the self-attention (see Fig.˜6) to preserve the facial attribute during pose customization to understand why the extension of the attention dimension helps to preserve the facial features. In order to do that, we took the generated images with CraftGraffiti and pass them through the self-attention layer of FLUX.1 dev, Subject-driven self-attention layer of Consistory (Tewel et al., 2024), and our face consistent self-attention layer. It has been observed that the traditional self-attention of FLUX.1 dev focuses on a certain point, whereas Subject-driven Self-Attention of ConsiStory (Tewel et al., 2024) attends the global features rather than the local ones. On the other hand, our proposed face-consistent self-attention, focused on the local features of the facial attributes, helps to maintain facial attributes during face customization.

Human Evaluation: We have conducted the perceptual study in order to evaluate the credibility and widespread adoption of the graffiti poster synthesized with CraftGraffiti. We chose 60 graffiti posters (20 for each pose) synthesized by (Ye et al., 2023; Labs et al., 2025; Brooks et al., 2023; Grok, 3; Hurst et al., 2024) and CraftGraffiti and asked the users to rate the graffiti on a scale of 1 to 5 (1 Low; 5 high) based on aesthetics, style blending, and recognizability. 47 anonymous users participated in this study. The outcome of this user study is reported in Fig.˜7.

For the user, CraftGraffiti makes a perfect graffiti-style poster, while other models either change the background or make it more cartoonish. In terms of recognizability CraftGraffiti preserves the local features (eyes, nose, lips, and so on) but changes the global features during style transfer, maintaining a decent recognizability compared to the rest.

5 Discussion on CraftGraffiti Performance

Technical Contributions and Performance: Our results demonstrate that integrating a face-consistent self-attention mechanism with LoRA-based fine-tuning allows CraftGraffiti to preserve key facial attributes across diverse poses and graffiti styles. Compared to existing pose-editing and style-transfer methods, CraftGraffiti achieves state-of-the-art identity preservation without compromising artistic quality. This balance between recognizability and stylization is essential for applications where personal identity and cultural authenticity are intertwined.

Cultural and Societal Implications: Beyond technical performance, CraftGraffiti acts as an enabler in ongoing cultural debates around representation, bias, and access to AI-generated art. By using graffiti, a medium historically tied to marginalized voices, we highlight the potential for generative systems to serve as inclusive cultural tools. However, these systems also risk reproducing and amplifying harmful biases, which must be addressed through dataset curation, fairness-aware training, and critical engagement with affected communities. By deploying the installation at Festival Cruïlla, we had an opportunity to observe user interactions in a high-energy, socially diverse, real-world context. This aligns with the principles of ecological validity (Kieffer, 2017) and living lab methodologies (8; 29). The festival setting facilitated spontaneous engagement, enabling us to capture authentic emotional responses and uncover interaction patterns that may not emerge in controlled lab settings (Pournaras et al., 2021). Such environments also pose technical challenges in terms of managing environmental noise, variable lighting, unpredictable crowd dynamics, and regulatory uncertainties. In our specific case, it appeared as an excellent example to dry test the new European AI Act. The implications of this regulatory learning approach are of enormous interest, though they are beyond the scope of this piece of work.

Limitations: While our model successfully preserves facial attributes, real-time generation in public settings is computationally demanding, and user interactions are influenced by the physical constraints of the installation space. Our evaluation was conducted in a single cultural context (a European music festival), which may limit the generalizability of the findings to other socio-cultural settings. However, general considerations can be taken regarding the questions posed: CraftGraffiti produces graffiti-styled human figures that preserve the original subject’s identity by maintaining key facial attributes across poses and styles through a face-consistent self-attention mechanism. Unlike many generative systems, it inherits body shape from the input image, avoiding homogenization into a single aesthetic norm. From an anthropological perspective, the tendency of many AI models to generate light-skinned, symmetrical, youthful, and slender bodies reflects the reinforcement of dominant, often Eurocentric, beauty standards (Muñoz et al., 2023). Such homogenization risks erasing diversity in age, body size, ability, and ethnic features, mirroring historical processes where global media displaced local aesthetics. Addressing these issues requires combining computational audits with ethnographic insights to prevent generative AI from fostering an aesthetic monoculture. Last but not least, it is of crucial importance to highlight the need for a strict control on the access of children, since, even though the interaction with the installation could be perceived as an entertainment action both from the children’s and parent’s perspective, it could create an actual unexpected impact in a moment of body development with psychological uncertainties.

Future Directions: Future work is definitly needed, and potential lines or work will have to explore: (1) extending CraftGraffiti to handle a broader range of cultural art forms beyond graffiti; (2) integrating real-time bias detection and mitigation pipelines; (3) testing in varied ecological environments, including community art events and educational settings; and (4) expanding cross-cultural studies to assess how identity preservation and stylistic adaptation are perceived in different cultural contexts.

6 Conclusion

CraftGraffiti represents a significant step forward in the intersection of generative AI and urban art, enabling the creation of personalized graffiti that preserves facial identity while embracing stylistic transformation. By introducing a face-consistent self-attention mechanism and leveraging LoRA-based fine-tuning, the system ensures that key facial features such as the eyes, nose, and mouth remain consistent across pose and style changes, addressing the longstanding challenge of identity preservation in diffusion-based image generation. This technical innovation allows for high-quality graffiti renderings that are both aesthetically compelling and faithful to the original subject, blending artistic stylization with personal recognizability. Beyond its technical contributions, CraftGraffiti holds broader societal and artistic implications by democratizing access to street art and empowering individuals and communities to participate in shaping their visual environments. By maintaining cultural motifs and individual likenesses, the model enables the creation of meaningful, identity-rich urban expressions that reflect diverse human experiences. As a result, CraftGraffiti not only advances the state of the art in generative customization but also offers a transformative tool for preserving and celebrating cultural narratives through the accessible medium of AI-generated graffiti.

Appendix A Demonstration at Cruïlla Festival

An installation (see Fig.˜8) implementing the proposed system was deployed during the Festival Cruïlla (www.cruillabarcelona.com) in Barcelona from 9-12 July 2025. The festival hosts 80,000 people during 4 days, attending parallel contexts and activities related to culture and innovation. More than 1,100 people visited the booth in which the installation was set, and 586 people were able to create their impression of a graffiti-style poster (see Fig.˜9) and give their feedback anonymously. We have summarized the human feedback and listed it as follows:

(1) The general impression of the users was that the demo was fun and engaging, often reacting with surprise and amusement at the results. The majority of participants understood that the demonstration was intended as a playful, exploratory experience rather than a precise or professional tool.

(2) Some users asked if their photos were being stored or used to train the algorithm, highlighting the importance of transparency in data management. We informed that none of the user’s data has been stored, referring to the informed consent regarding the use of the images only for quality control. The output images of Fig.˜9 are are obtained with the permission of the user to showcase the output of the generation for solely research purpose.

(3)The system tended to apply heavy makeup to women’s faces, regardless of whether the user was wearing any in the original photo. Generated images of women occasionally included exaggerated features, such as full lips and prominent cheekbones.

(4) A large number of participants commented that the results looked quite similar and their identity was reflected by the system. This was also consistent when asked about their opinion on how the system represented other participants, across users, with only slight variations in hairstyle or clothing, reducing the perceived uniqueness.

(5) The system tended to make people look younger, except for the younger people, in which an explicit evolution on the perceived maturity was identified, particularly in aspects related to body shape, such as muscular development for men or stylized forms for women.

(6) Even though gender consistency is high, for a limited number of cases, some women subjects were interpreted as men. In this case, the user was immediately addressed by the installation operators to validate the reaction and re-state the experimental nature of the installation, addressing the particular misfunctions that it can have and providing valuable feedback for the improvement.

Appendix B Some more qualitative examples

We generated some more qualitative examples with CraftGraffiti as reported in Fig.˜10. It has been observed that, if the user uses some external facial accessories (e.g., classes) CraftGraffiti also preserves it in graffiti style (1st and 2nd row of Fig.˜10). Similarly, if the user is bald or wearing hijab, it maintains that hair pattern and the external clothing accessories too (3rd and 4th row of Fig.˜10).

Appendix C Future Directions

Future work is definitly needed, and potential lines or work will have to explore: (1) extending CraftGraffiti to handle a broader range of cultural art forms beyond graffiti; (2) integrating real-time bias detection and mitigation pipelines; (3) testing in varied ecological environments, including community art events and educational settings; and (4) expanding cross-cultural studies to assess how identity preservation and stylistic adaptation are perceived in different cultural contexts.

Appendix D Ethics Statement

The application of CraftGraffiti in image editing offers extensive potential for diverse downstream applications, facilitating the adaptation of images to different contexts. The primary objective of our model is to automate and streamline this process, resulting in significant time and resource savings. It is important to acknowledge that current methods have inherent limitations, as discussed in this paper. However, our model can serve as an intermediary solution, expediting the creation process and offering valuable insights for further advancements. It is crucial to remain mindful of potential risks associated with these models, including the dissemination of misinformation, potential for abuse, and introduction of biases. Broader impacts and ethical considerations should be thoroughly addressed and studied in order to responsibly harness the capabilities of such models. Terms and conditions forms were explicitly offered to all the participants during the live festival, and researchers and support team were always present during the realization of the images, vigilant to potential issues appearing and aware of the user’s reactions.

Appendix E Code for facial consistency evaluation

Through the following implementation, we first load the FaceNet [43] model pretrained on VGGFace2 [6], which is designed to produce numerical vector embeddings for faces. Given two embedding vectors

[TABLE]

the cosine similarity is defined as:

[TABLE]

where:

• •

$\mathbf{u}\cdot\mathbf{v}$ is the dot product of the two vectors.

• •

$\|\mathbf{u}\|_{2}$ and $\|\mathbf{v}\|_{2}$ are their Euclidean norms.

A value close to $1$ indicates that the two embeddings are highly similar in direction, meaning the corresponding faces are likely to be of the same identity. A value near [math] indicates no similarity, while negative values imply opposite feature orientations.