Replication for Characterizing Photorealism (2025 Paper)

This repository contains the replication materials for the "Characterizing Photorealism" paper. The repository includes the necessary code and instructions to replicate the analysis presented in the paper.

Data and Stimuli

• The stimuli is available on Zenodo.

• Open and run jupyter notebook CHI2025_feb.ipynb

Data Files

This repository contains several CSV files used in the analysis. Below is a description of each:

1. Image Metadata (image_metadata.csv)

• Contains filenames of all images used in the experiment.
• Indicates whether each image is real or AI-generated.

2. User Guesses (data.csv)

• Contains participant responses from the online experiment.
• Includes 749,828 user entries with details such as:

3. Artifact Annotations (artifact_annotations.csv)

• Contains images and their associated artifact annotations.
• Includes a description of each artifact type and the reasoning behind the classification.

4. Human Curation (human_curation.csv)

• Evaluates the effect of human curation on image accuracy.

5. Classified Comments (classified_comments_chunked.csv)

• Contains participant comments collected from the online experiment.
• Each comment is classified into one or more artifact categories based on the themes identified by GPT-3.5 Turbo.

If you use this dataset or code, please cite:

@inproceedings{kamali2025photorealism, author = {Negar Kamali and Karyn Nakamura and Aakriti Kumar and Angelos Chatzimparmpas and Jessica Hullman and Matthew Groh}, title = {Characterizing Photorealism and Artifacts in Diffusion Model-Generated Images}, booktitle = {CHI Conference on Human Factors in Computing Systems (CHI '25)}, year = {2025}, publisher = {Association for Computing Machinery}, address = {Yokohama, Japan}, doi = {10.1145/3706598.3713962}, isbn = {979-8-4007-1394-1}, keywords = {photorealism, diffusion models, generative AI, synthetic media, deepfakes, misinformation} }