---
license: mit
tags:
- deepfake-detection
- ai-generated-image-detection
- synthetic-media
- image-forensics
- computer-vision
- clip
- siglip
base_model:
- openai/clip-vit-large-patch14
- google/siglip2-base-patch16-512
pipeline_tag: image-classification
---

![DeepReality](https://huggingface.co/OmerKurtulus/deepreality-models/resolve/main/cover.png)

# DeepReality: Layer 2 detector weights

Trained weights for the Layer 2 detection core of **DeepReality**, a six-layer
forensic system for detecting AI-generated and manipulated imagery.

**Source code, documentation and the full pipeline:**
https://github.com/OmerKurtulus/DeepReality

These files are the three detectors that were trained for the project. They are
not a standalone classifier: DeepReality reads them as three of fifteen
independent analysis modules, and its verdict is produced by a reasoning layer
that weighs them against documentary provenance, compression physics and
localisation evidence. Used alone, they carry the limitations described below.

## Files

| File | Pin | Size | Architecture |
|---|---|---|---|
| `pin_b1_clip_ln_tune_final.pt` | PIN-B1 | 1.6 GB | CLIP ViT-L/14, frozen backbone with LayerNorm tuning |
| `pin_b2_siglip2_finetune_final.pt` | PIN-B2 | 1.4 GB | SigLIP2-base-patch16-512, full fine-tune |
| `pin_b3_freq_cnn_final.pt` | PIN-B3 | 19 MB | Frequency CNN over DCT/DWT maps, trained from scratch |

Two further models complete the detection core and are **not** hosted here.
PIN-B4 is a public third-party checkpoint
([prithivMLmods/AI-vs-Deepfake-vs-Real-Siglip2](https://huggingface.co/prithivMLmods/AI-vs-Deepfake-vs-Real-Siglip2)),
and the PIN-F1 XGBoost meta-learner is small enough to be tracked directly in
the GitHub repository.

## Usage

Download the three files into the `models/` directory of a DeepReality clone:

```bash
git clone https://github.com/OmerKurtulus/DeepReality.git
cd DeepReality

hf download OmerKurtulus/deepreality-models --local-dir models
```

PIN-B1 and PIN-B2 additionally fetch their base architectures from the Hub on
first run. Placement instructions and the remaining setup are in
[`models/README.md`](https://github.com/OmerKurtulus/DeepReality/blob/main/models/README.md).

## The system these weights belong to

A single image taken through all fifteen pins. Every figure is the output of one
real pipeline run, not an illustration. PIN-B1, PIN-B2 and PIN-B3 are the three
detectors hosted in this repository; the verdict is not theirs to make.

![Worked example](https://huggingface.co/OmerKurtulus/deepreality-models/resolve/main/preview.png)

Note what happens in that run. The four detectors agree at a mean of 0.932, and
the reasoning layer still decides on the C2PA manifest instead, lowers its own
confidence because the signing certificate had expired, and flags that three of
the four detectors share a training corpus, so their agreement is correlated
rather than independent.

## Training data

All three detectors were trained on **OpenDeepfake-Preview**, 20,000 images.

## Reported performance

Held-out results on the training corpus:

| Pin | Accuracy | ROC-AUC |
|---|---|---|
| PIN-B1 | 99.77% | 0.9997 |
| PIN-B2 | 99.97% | 1.0000 |
| PIN-B3 | 96.50% | 0.9923 |

## Limitations, and why they matter

**Read the figures above as in-distribution performance. They do not survive a
change of corpus.** Measured on
[`ComplexDataLab/OpenFake`](https://huggingface.co/datasets/ComplexDataLab/OpenFake)
the same detectors fall to 0.51 to 0.84 ROC-AUC. On
[`Hemg/deepfake-and-real-images`](https://huggingface.co/datasets/Hemg/deepfake-and-real-images)
all four sit at chance, and PIN-B3 is anti-correlated with the label, which is
worse than noise because a downstream fusion stage learns to trust it backwards.

This generalisation gap is the reason DeepReality exists. A single detector
reporting 99% on its own benchmark tells you very little about an unseen
generator, which is why the system combines paradigms that fail differently and
subordinates all of them to documentary provenance when provenance is present.

The only figure measured by cross-validation on a corpus none of these detectors
saw in training is the Layer 6 meta-learner: ROC-AUC 0.8846 at an Expected
Calibration Error of 0.0203. That is the conservative number, and the one to
quote.

**Intended use.** Research and assistive review. These weights should not be
used as the sole basis for a consequential decision about a specific image or
person.

## Licence and citation

Released under the MIT licence. If you use this work, please cite it:

```bibtex
@software{kurtulus2026deepreality,
  author  = {Kurtulus, Omer Faruk},
  title   = {DeepReality: PIN Architecture for Multi-Layer AI Image Forensics},
  year    = {2026},
  url     = {https://github.com/OmerKurtulus/DeepReality},
  license = {MIT}
}
```