GPTTrace image detector: Community Forensics ViT-S fine-tune v1 (int8 ONNX)
This model detects AI-generated images. It is GPTTrace's fine-tune of the Community Forensics ViT-S detector on newer generators. On gpttrace.com it runs in the browser next to the base model, and their logits are averaged.
Base: Community Forensics ViT-S (MIT), trained on 2.7M images from 4,803 generators
Fine-tuned on commercially usable data:
- DRAGON (CC BY-SA 4.0)
- open-image-preferences (Apache-2.0)
- bitmind nano-banana (MIT)
- megalith CC0 photos
Every AI image is re-encoded as JPEG or WebP during training.
Licence: CC BY-SA 4.0. The DRAGON images are share-alike, so we release the fine-tuned weights under the same licence.
File:
onnx/model_int8.onnx(int8 on MatMul only), 24 MB
Measured results
| Test | Base | Average of base + this model |
|---|---|---|
| AIGIBench (NC test set, evaluation only), centre crop | AUC 0.949 | — |
| AIGIBench, max of centre and native-resolution crops | 0.979 | AUC 0.993 |
| OpenFake (NC test set, evaluation only) | — | AUC 0.960 |
| DRAGON + megalith held-out | 0.997 | 1.000 |
| GPTTrace in-house set, full scorer, 5-fold CV | — | AUC 0.980; 84% of AI images caught at 5% false positives on real photos |
Known weak spots: FLUX dev (43% caught) and GPT Image (53% caught) on our in-house set.
Preprocessing
- Resize so the shortest edge is 440.
- Centre-crop to 384.
- Normalise with mean (0.4815, 0.4578, 0.4082) and std (0.2686, 0.2613, 0.2758). Input is NCHW float32.
The output is one logit; higher means AI.
Training and evaluation code: github.com/andrewnakas/gpttrace (eval/).