---
license: mit
tags:
  - deepfake-detection
  - ai-image-detection
  - dfd-arena
  - bitmind
library_name: transformers
pipeline_tag: image-classification
---

# PixelPrism v0.1 — DFD Arena Submission

Sanitized single-detector submission for the [BitMind Deepfake Detection Arena](https://huggingface.co/spaces/bitmind/dfd-arena-leaderboard).

This repo represents the **most informative single component** of PixelPrism's
production [V9 16-detector ensemble](https://pixelprism.ai/leaderboard), wrapped
in the BitMind `DeepfakeDetector` interface so it can be evaluated alongside
NPR / UCF / CAMO on the public leaderboard.

## What's in this submission

A wrapper around the **Swin V2 transformer head**
([haywoodsloan/ai-image-detector-deploy](https://huggingface.co/haywoodsloan/ai-image-detector-deploy),
MIT licensed). In PixelPrism's V9 permutation-importance audit (8000 samples,
5 reps), Swin V2 ranked **#1 by a wide margin** at importance **0.271**, vs
0.109 for the next-best detector (vit3) and 0.032 for DIRE-FLUX. It alone
accounts for ~38% of V9's total discriminative power.

## What's NOT in this submission

The full PixelPrism V9 ensemble fuses **16 detectors** via a
`HistGradientBoostingClassifier` meta-classifier:

```
fft, vit, vit2, vit3, dire (SD 1.5), clip, srm, exif, face,
cfa, prnu, c2pa, anatomy, swin, dire_sdxl, dire_flux
```

Some V9 components depend on weights that are not MIT-redistributable:
- `dire_flux` uses FLUX.1-schnell (non-commercial license)
- `dire_sdxl` uses Stability SDXL (CreativeML OpenRAIL-M)
- `face` uses FaceForensics++ Xception variants (access-gated)

Those stay in our internal production stack rather than the public submission.

## Live full-ensemble numbers

The full V9 ensemble is live at **<https://pixelprism.ai/api/detect>** (paid)
and **<https://pixelprism.ai/api/scan-public>** (5/day free tier). Per-generator
detection rates and 30-day drift trend are published at
**<https://pixelprism.ai/leaderboard>** (refreshed monthly with each retrain).

V9 internal holdout (8000 stratified samples, 4000 real / 4000 AI):

| Metric | V9 |
|---|---|
| Overall | 96.7% |
| Real | 96.1% |
| AI | 97.4% |
| Per-generator min | 91.0% (Grok) |
| Drift gap (fresh AI vs known AI) | −2.7pp (fresh AI now BEATS known AI) |

## Files in this repo

| File | Purpose |
|---|---|
| `pixelprism_detector.py` | The `DeepfakeDetector` subclass registered as `PixelPrism` in `DETECTOR_REGISTRY` |
| `pixelprism_config.yaml` | YAML config with `hf_repo`, `backbone_repo`, `ai_label_idx` |
| `model.safetensors` | Swin V2 weights (re-hosted, byte-identical to upstream) |
| `config.json` | Swin V2 model config |
| `preprocessor_config.json` | Swin V2 image preprocessor config |
| `README.md` | This file |

## Citation / contact

If you use this in research or a comparison study, cite:
- The Swin V2 detector: [haywoodsloan/ai-image-detector-deploy](https://huggingface.co/haywoodsloan/ai-image-detector-deploy)
- PixelPrism's full ensemble methodology: <https://pixelprism.ai/leaderboard>

Operator: Chris Crawley, PixelPrism.ai · <https://pixelprism.ai>

## License

MIT (matches the upstream Swin V2 model + matches the BitMind DFD Arena requirement).