PICTOS-AI: Generating Cognitively Accessible Pictograms with Artificial Intelligence for Inclusive Visual Communication
| Título | PICTOS-AI: Generating Cognitively Accessible Pictograms with Artificial Intelligence for Inclusive Visual Communication |
|---|---|
| Año | 2025 |
| Autor | Gabriel Olmos Leiva, Ashley Jara Goicovich, Herbert Spencer González, Gabriel Hermosilla Vigneau |
| Tipo | Ponencia, Proceeding |
| Editorial | IEEE |
| Edición | Chilecon 2025 |
| Ciudad | Valparaíso |
| Páginas | 1-7 |
| Palabras Clave | Generative AI, DreamBooth, LoRA, fine-tuning, Pictogram generation, Flow-Matching, Accessible communication, Cognitive accessibility |
| Área de Investigación | Forma, Cultura y Tecnología |
| Código | 10.1109/CHILECON66915.2025.11476466 |
| URL | https://ieeexplore.ieee.org/document/11476466 |
| Carreras | Diseño, Otra |
| Nota | Explores the use of generative AI to automate the creation of pictograms coherent with the PICTOS style for inclusive visual communication. PICTOS project proposes pictographic system structured in three layers: action, element, and context facilitating procedure understanding. Applies fine-tuning using DreamBooth and LoRA techniques to Flux1-dev diffusion model, training with a dataset structured according to the PICTOS system logic. Presents specialized training methodology along with exhaustive model evaluation using visual fidelity, perceptual diversity, and subject fidelity metrics. Results show fine-tuning allows generation of stylistically consistent and semantically relevant pictograms for public service domain. |
Citation
- G. O. Leiva, A. J. Goicovich, H. S. González and G. H. Vigneau, "PICTOS-AI: Generating Cognitively Accessible Pictograms with Artificial Intelligence for Inclusive Visual Communication," 2025 IEEE CHILEAN Conference on Electrical, Electronics Engineering, Information and Communication Technologies (CHILECON), Valparaíso, Chile, 2025, pp. 1-7, doi: 10.1109/CHILECON66915.2025.11476466.
Abstract
1.3 billion people worldwide 16% of global population) experience significant disabilities, with 13.9% of adults facing cognitive impairments that create barriers to accessing urban services. This work presents PICTOS-AI, the first application of flow-matching diffusion models to automated generation of cognitively accessible pictograms. We fine-tune the Flux.1dev model using DreamBooth and LoRA techniques with only 25 training images, achieving remarkable improvements: 53% enhancement in perceptual similarity (LPIPS: $0.627 \rightarrow 0.295$) and 101% increase in subject fidelity (DINO: $0. 3 4 3 \rightarrow 0. 6 8 9$). Our approach maintains stylistic consistency with the three-layer PICTOS system while enabling scalable generation of accessible visual communication. Results demonstrate the technical feasibility of transforming manual pictogram creation into an automated solution, representing a significant step toward more inclusive urban communication systems that could potentially benefit populations facing cognitive and linguistic accessibility barriers.