TactileNet: Bridging the Accessibility Gap with AI-Generated Tactile Graphics for Individuals with Vision Impairment

TL;DR

This work tackles the critical need for scalable tactile graphics for visually impaired learners by proposing TactileNet, a domain-specific dataset (1,029 tactile images across 66 classes) and an AI-driven pipeline that generates embossing-ready 2D tactile templates using text-to-image Stable Diffusion fine-tuned with LoRA and DreamBooth. The approach combines a 66-class adapter training strategy with prompt-based editing, enabling both text-only and image-conditioned generation, and is evaluated through a web-based, expert-driven protocol showing high adherence to tactile guidelines ($0.<q>$) and near-human structural fidelity (SSIM around $0.538$–$0.549$). Key findings show strong pose alignment and meaningful silhouette preservation, with generated outputs closely mirroring expert designs while enabling customization (e.g., removing logo or pocket details). The work demonstrates a practical, scalable path to AI-augmented tactile design that accelerates production without replacing human expertise, and commits to publicly releasing code, data, and models to foster further research in accessible education.

Abstract

Tactile graphics are essential for providing access to visual information for the 43 million people globally living with vision loss. Traditional methods for creating these graphics are labor-intensive and cannot meet growing demand. We introduce TactileNet, the first comprehensive dataset and AI-driven framework for generating embossing-ready 2D tactile templates using text-to-image Stable Diffusion (SD) models. By integrating Low-Rank Adaptation (LoRA) and DreamBooth, our method fine-tunes SD models to produce high-fidelity, guideline-compliant graphics while reducing computational costs. Quantitative evaluations with tactile experts show 92.86% adherence to accessibility standards. Structural fidelity analysis revealed near-human design similarity, with an SSIM of 0.538 between generated graphics and expert-designed tactile images. Notably, our method preserves object silhouettes better than human designs (SSIM = 0.259 vs. 0.215 for binary masks), addressing a key limitation of manual tactile abstraction. The framework scales to 32,000 images (7,050 high-quality) across 66 classes, with prompt editing enabling customizable outputs (e.g., adding or removing details). By automating the 2D template generation step-compatible with standard embossing workflows-TactileNet accelerates production while preserving design flexibility. This work demonstrates how AI can augment (not replace) human expertise to bridge the accessibility gap in education and beyond. Code, data, and models will be publicly released to foster further research.

Figures (8)

• Figure 1: Example tactile graphics of a horse. Left: Expert-designed tactile graphic from existing libraries. Right: Tactile graphic generated by our fine-tuned model.
• Figure 2: Examples of our image-to-image translation framework: Top row shows reference natural images, middle row displays TactileNet's sourced/benchmark tactile graphics, and the bottom row presents generated tactile graphics using our adapters.
• Figure 3: Flow diagram illustrating the process of data compilation from initial sourcing to final dataset compilation.
• Figure 4: Comprehensive workflow of our framework, starting with fine-tuning (left), where TactileNet data (tactile images, text prompts) refine the SD model. The process transitions to the generation phase (right), applying fine-tuned adapters atop the frozen SD model for text-to-image and image-to-image tactile graphic generation.
• Figure 5: Failure cases from the instruction-tuned baseline. Although the test-time natural images (top row) are diverse and unseen during training, the model fails to adapt and instead memorizes training prompt-class pairs, producing incorrect tactile outputs (bottom row). Memorized outputs correspond to tomato, helicopter, and bed classes.