# PairCLIP-SWA reports stronger deepfake detection with one model

> A study finds that averaging a detector’s training checkpoints improved results on two image benchmarks, including a case where validation rankings misled model selection.

Oossa · 2026-10-30 · https://oossa.com/en/pairclip-swa-reports-stronger-deepfake-detection-with-one-model

Researchers introduced PairCLIP-SWA, a deepfake image detector built on one OpenCLIP vision model. On the IEEE SP Cup 2025 benchmark, they report 97.04% accuracy, compared with 94.46% for a published three-model ensemble trained under the same protocol.

The researchers also found that checkpoint rankings on validation data could be misleading: on an OpenForensics-derived test set, the validation-selected checkpoint scored 88.85%, while the averaged model scored 94.76%. These are results from the paper’s experiments, not evidence of performance on every kind of deepfake.

## The facts

- On IEEE SP Cup 2025, PairCLIP-SWA scored 97.04% accuracy and 98.64% AU-ROC, according to the researchers.
- On a 10,905-image OpenForensics-derived test partition, the validation-selected checkpoint scored 88.85%; the terminal average scored 94.76%.

## Why it matters

For teams choosing a deepfake detector, the results suggest that relying on the single checkpoint with the best validation score can lead to weaker performance on separate test images; performance beyond these benchmarks remains unknown.

## Sources & references

1. [PairCLIP-SWA: Similarity-Guided Balancing and Stochastic Weight Averaging for Generalized Deepfake Image Detection with OpenCLIP](https://arxiv.org/abs/2610.37874) – arXiv cs.CV, 2026-10-30

Last updated: 2026-10-30