sdxl-detector

SDXL Detector

This model was created by fine-tuning the umm-maybe AI art detector on a dataset of Wikimedia-SDXL image pairs, where the SDXL image is generated using a prompt based upon a BLIP-generated caption describing the Wikimedia image.

This model demonstrates greatly improved performance over the umm-maybe detector on images generated by more recent diffusion models as well as non-artistic imagery (given the broader range of subjects depicted in the random sample drawn from Wikimedia).

However, its performance may be lower for images generated using models other than SDXL. In particular, this model underperforms the original detector for images generated using older models (such as VQGAN+CLIP).

The data used for this fine-tune is either synthetic (generated by SDXL) and therefore non-copyrightable, or downloaded from Wikimedia and therefore meeting their definition of "free data" (see https://commons.wikimedia.org/wiki/Commons:Licensing for details). However, the original umm-maybe AI art detector was trained on data scraped from image links in Reddit posts, some of which may be copyrighted. Therefore this model as well as its predecessor should be considered appropriate for non-commercial (i.e. personal or educational) fair uses only.

Model Trained Using AutoTrain

• Problem type: Image Classification

Validation Metrics

loss: 0.08717025071382523

f1: 0.9732620320855615

precision: 0.994535519125683

recall: 0.9528795811518325

auc: 0.9980461893059392

accuracy: 0.9812734082397003