Authors: Joshua Peterson, Stefan Uddenberg, Thomas Griffiths, Alexander Todorov, Jordan Suchow

The One Million Impressions (OMI) dataset contains over a million human judgments of over 1,000 synthetic face images along 34 attributes.

Please note that these attribute inferences, especially those of the more subjective or socially constructed attributes, have no necessary correspondence to the actual identities, attitudes, or competencies of people whom the images resemble or depict (e.g., a trustworthy person may be wrongly assumed to be untrustworthy on the basis of appearance). Rather, these inferences, and in turn our measurements, reflect systematic biases and stereotypes about attributes shared by the population of raters. If you are considering studying or applying this dataset in your own work, please be mindful that failing to make this critical distinction can result in perpetuating the biases studied here.

When making use of the data in your research, please cite the associated publication:

Peterson, J. C., Uddenberg, S., Griffiths, T. L., Todorov, A., & Suchow, J. W. (2022). Deep models of superficial face judgments. Proceedings of the National Academy of Sciences (PNAS), 119(17), e2115228119. doi:10.1073/pnas.2115228119

BibTeX entry:

@article{peterson2022omi,
  title={Deep models of superficial face judgments},
  author={Peterson, Joshua C and Uddenberg, Stefan and Griffiths, 
  Thomas L and Todorov, Alex and Suchow, Jordan W},
  journal={Proceedings of the National Academy of Sciences},
  volume = {119},
  number = {17},
  pages = {e2115228119},
  year={2022},
  doi = {10.1073/pnas.2115228119},
  publisher={National Acad Sciences},
  URL = {https://www.pnas.org/doi/abs/10.1073/pnas.2115228119},
}

The data is distributed under the Creative Commons BY-NC-SA 4.0 license. If you intend to use it, please see LICENSE.txt for more information.

Please make sure to check the main repository link for any updated or corrected versions of this dataset.

Mean judgements on each of the 34 attributes for each of the 1,004 face stimuli can be found in attribute_means.csv. The corresponding synthetic face stimuli are located in the images folder. The original, raw approximately 1.3 million individual judgments can be found in attribute_ratings.csv in the attribute_ratings.zip archive.

We also include data from the additional validation experiments reported in Peterson et al. (2022). The aggregated data is provided in processed_validation_data.csv, and the more fine-grained trial-level data is provided in raw_validation_data.csv, the legend for which is below:

• rt = how long it took for the participant to complete the trial
• condition = attribute we manipulated the faces along
• response = the rating from 0 to 100 that the participant gave for that face
• stimulus_type = real or artificial face
• stimulus_index = the ID of the face. for those from the Chicago Face Dataset, they start with "CFD", while artificial faces are some arbitrary number greater than 1004
• stimulus_level = 0, 1, or 2 for low, medium, and high levels of that attribute. In practice 1 is mean, 0 is -0.5 SD, and 2 is +0.5 SD
• repeat = whether this is a repeat showing of the face

The face image stimuli for these experiments are located in the validation_images folder.

The face dataset used to create underlying model components that in turn were used to generate our synthetic faces can be found at: https://github.com/NVlabs/ffhq-dataset. The model components themselves can be found at: https://github.com/NVlabs/stylegan2