bodhisattamaiti/EmotionVis

EmotionVis is a structured prompt benchmark designed for emotion-conditioned visual generation. The dataset consists of 1,000 carefully constructed prompts that combine diverse environments with a range of emotional attributes, enabling controlled exploration of how generative models interpret and render emotional context in visual scenes. 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) emotion-specific cues (e.g., soft light for peaceful, dark clouds for tense). These components are combined into a final prompt that is designed to be expressive, consistent, and suitable for high-quality image generation across a wide range of generative models. The dataset is intentionally model-agnostic and does not include any generation-specific parameters such as seeds, resolution, or inference steps. This allows users to apply the prompts across different frameworks, including diffusion models (e.g., SDXL), autoregressive image models, or proprietary systems, without constraints. To support controlled experimentation, each environment–emotion pair is associated with multiple guidance-scale values, allowing researchers to study the effect of prompt adherence and stylistic intensity under varying levels of conditioning strength.

🎯 Intended Use EmotionVis is designed to support: Evaluation of emotion-conditioned image generation Prompt engineering and prompt sensitivity analysis Cross-model comparison of emotional expressiveness Studies on alignment between textual emotion cues and visual outputs Benchmarking robustness across different generation settings The dataset can also serve as a foundation for building extended benchmarks involving style variation, object-centric augmentation, or multimodal evaluation.

⚠️ Limitations The dataset contains only textual prompts and does not include generated images. Emotional interpretation is inherently subjective and may vary across models and evaluators. The prompts are designed to avoid human subjects to ensure reliable rendering across models, which may limit certain types of emotional expression.

🔄 Future Extensions EmotionVis is designed to be extensible. Potential future directions include: Expansion to additional emotions and environments Inclusion of object-centric emotional prompts Style-conditioned prompt variants Human evaluation annotations Multimodal extensions (e.g., audio or textual narratives)

🤝 Contributors

• Debshree Chowdhury, Bodhisatta Maiti

📝 Citation @dataset{emotionvis, title = {EmotionVis: A Prompt Benchmark for Emotion-Conditioned Visual Generation}, author = {Debshree Chowdhury and Bodhisatta Maiti}, year = {2026}, note = {Hugging Face Dataset}, }