Creating and validating photorealistic AI-generated facial expression stimuli for emotion research

The development of AI-generated facial expression stimuli offers researchers unprecedented control over stimulus characteristics while maintaining photorealistic quality. However, rigorous validation methodologies are essential to ensure these stimuli function effectively in behavioral research. This paper presents a comprehensive three-study validation framework for AI-generated facial expression stimuli that systematically varied body weight and emotional expressions. Study 1 validated 72 stimuli across three emotions (happiness, sadness, neutrality) and three weight categories using both computational and human validation approaches. Study 2 extended this framework to include anger as a fourth emotion while controlling for individual facial features through same-character methodology. Study 2b further assessed valence, intensity, naturalness, authenticity, and non-focal emotion ratings of the same stimuli from Study 2. Across studies, the stimuli demonstrated successful emotion and weight manipulations, with AI-generated images achieving photorealistic quality that participants could not reliably distinguish from real photographs. Computational validation showed high intended-emotion likelihood in Py-Feat, while human validation confirmed appropriate perception of the target emotions, body weight, realism, age, and broader expression-quality dimensions. This dual validation approach-combining automated computational assessment with human behavioral validation-provides a robust framework for ensuring AI-generated stimuli meet research standards. The methodology offers significant advantages for behavioral research, enabling precise control over stimulus characteristics while maintaining ecological validity and addressing ethical considerations in stimulus development.