bodhisattamaiti/SoundSensitivityVis

SoundSensitivityVis is a structured prompt benchmark designed to study how generative models translate acoustic attributes into visual representations. The dataset consists of 1,000 carefully constructed prompts that combine diverse environments with abstract sound profiles, enabling controlled exploration of sound-to-visual mapping in image generation systems. Each prompt is composed of three core components: (1) a base environment (e.g., forest trail, lighthouse coast), (2) environment-specific contextual details (e.g., dense foliage, rocky shoreline), and (3) sound-derived visual cues (e.g., heavy atmosphere for low-frequency rumble, sharp contrast for high-pitched signals). These components are combined into a final prompt that is designed to be expressive, consistent, and compatible with a wide range of generative models. The dataset defines sound in terms of abstract acoustic properties—such as loudness, frequency characteristics, temporal patterns, and texture—rather than specific sound sources. This abstraction allows the dataset to focus on property-level cross-modal alignment, where acoustic features are translated into corresponding visual attributes like lighting, contrast, composition, and texture. To support controlled experimentation, each environment–sound profile pair is associated with multiple guidance-scale values, enabling the study of how varying levels of prompt conditioning influence the visual manifestation of sound-related cues. SoundSensitivityVis is intentionally model-agnostic and does not include generation-specific parameters such as seeds, resolution, or inference steps. This allows users to apply the prompts across different generative frameworks without constraints.

🎯 Intended Use SoundSensitivityVis is designed to support: Evaluation of sound-conditioned visual generation Cross-modal alignment studies between acoustic and visual domains Prompt engineering and sensitivity analysis Comparative evaluation across generative models Analysis of how abstract sensory attributes influence visual outputs The dataset can also serve as a foundation for multimodal research involving audio-to-image generation, as well as combined emotion and sound conditioning.

⚠️ Limitations The dataset contains only textual prompts and does not include generated images. Sound profiles are represented abstractly and may be interpreted differently by different models. The prompts avoid human subjects to ensure consistent rendering across models, which may limit certain expressive scenarios.

🔄 Future Extensions SoundSensitivityVis is designed to be extensible. Potential future directions include: Inclusion of real-world sound sources and audio clips Expansion to additional acoustic dimensions and profiles Integration with emotion-conditioned prompts (e.g., combining EmotionVis and SoundSensitivityVis) Human evaluation annotations for perceptual alignment Multimodal datasets linking audio, text, and generated images

🤝 Contributors

• Debshree Chowdhury, Bodhisatta Maiti

📝 Citation @dataset{soundsensitivityvis, title = {SoundSensitivityVis: A Prompt Benchmark for Visual Response to Acoustic Attributes}, author = {Debshree Chowdhury and Bodhisatta Maiti}, year = {2026}, note = {Hugging Face Dataset}, }