Training a detector to read AI image fingerprints
TikTok TechJam 2026 · Track 5 · DINOv3 ViT-L
The budget is two billion parameters. We used 302 million: DINOv3 ViT-L, fully fine-tuned, with a global head for the page-level verdict and a local head for a 32×32 patch map.
Training is public data only, 2.58 million images, comprising AI generated images from different generator architectures — ranging from older GANs to frontier diffusion models. Analyze runs the checkpoint on an image you bring.
Architecture
The DINOv3 ViT-L/16 backbone is fine-tuned not just for image classification but also with patch-level supervision.
We put a heatmap on every result so the verdict is not a black box: it shows which patches the model treats as generated.
Held-out numbers
These are Seer's own held-out scores from various community datasets. The NTIRE 2026 open-test leaderboard sits under the table so the public-test AUROC can be read next to the published entries.
NTIRE 2026 open test
The open-test leaderboard of the NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild, held at CVPR 2026. ROC AUC and robust ROC AUC, from Table 3 of the NTIRE 2026 report. The published entries are 7B models; Seer is 302M and still third on robust AUROC — state of the art at this scale.
| Method | Params | ROC AUC | Rob. ROC AUC |
|---|---|---|---|
| MICV | 7B | 0.9978 | 0.9738 |
| Ant International | 7B | 0.9973 | 0.9731 |
| Seer (ours) | 302M | 0.9677 | 0.9228 |
| TeleAI-TeleGuard | 7B | 0.9762 | 0.9215 |
| INTSIG | 7B | 0.9810 | 0.9090 |
| vincentlc | 7B | 0.9497 | 0.8633 |
| UESTC | 7B | 0.9693 | 0.8558 |
| Reagvis Labs | 7B | 0.9423 | 0.8474 |
| PSU | 7B | 0.9132 | 0.8334 |
| Shallow Real | 7B | 0.9954 | 0.8302 |
Published scores from Table 3 of Gushchin et al., NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild
Held-out breakdowns
Where the misses live — by family, generator, and source.
The mixture
Ten public sources, weighted by measured difficulty. 2.58 million usable images: 1.70 million fake and 875 thousand real. We open a source for what is in it and how to fetch it. Weights decide draw probability, not disk usage.
usable by the mix: 2,576,437 images — 1,701,288 fake · 875,149 real.
What we trained with
The tools, models, libraries, and data that produced the numbers above. Only the parts that shaped the result.
Environment
40 hours of training on 16 vCPU (Xeon Platinum 8462Y+), 251 GB RAM, 30 GB container disk, and a 2 TB network volume at /workspace holding the 2.5M-image mixture.