A Pilot Study on Curator-Guided Multilingual Art Description for Blind and Low-Vision Audiences with Small Vision-Language Models

Researchers conducted a pilot study using small vision-language models (Qwen2.5-VL-3B-Instruct) to generate multilingual art descriptions for blind and low-vision audiences in museum settings. The study compared language-specific and multilingual adapter approaches across German, Romanian, and Serbian, finding that language-specific models performed better for accessibility while maintaining privacy through on-premise deployment.

This research addresses a significant accessibility gap in cultural institutions where blind and low-vision audiences struggle to engage with visual art, particularly across non-English languages. The study's use of small, on-premise vision-language models reflects growing institutional concerns about data privacy and intellectual property in museum contexts, where cloud-based AI services may be unsuitable. By deploying a 3-billion parameter model rather than larger alternatives, the researchers demonstrate that meaningful accessibility improvements don't require extensive computational resources.

The pilot's comparison of language-specific versus multilingual adapters yields nuanced findings: language-specific LoRA adapters provide more controllable and visually grounded descriptions for Romanian and Serbian, while multilingual approaches remain competitive for German. This suggests that accessibility effectiveness varies by language and cultural context rather than following universal patterns. The inclusion of an LLM-as-Judge evaluation protocol calibrated against actual BLV user feedback strengthens the methodology beyond typical benchmark metrics.

The findings carry implications for cultural institutions and AI practitioners. Museums can implement accessible AI systems locally without external vendor dependencies, reducing privacy risks while serving underserved audiences. However, the researchers deliberately frame their conclusions cautiously, acknowledging the pilot's limited scope and calling for larger, longitudinal studies with actual BLV users across broader language coverage. This measured approach prevents overgeneralization while establishing a baseline for future accessibility-focused AI research in cultural contexts.

• →Small on-premise vision-language models enable privacy-preserving accessibility solutions for museums without cloud dependency
• →Language-specific adapters outperformed multilingual approaches for Romanian and Serbian art descriptions in accessibility metrics
• →BLV user involvement in evaluation methodology proved essential for validating AI-generated descriptions beyond automated metrics
• →Curator-guided adaptation frameworks allow cultural institutions to control description quality and relevance for their collections
• →Further research with larger BLV cohorts and additional languages is required before establishing generalizable accessibility guidelines