---
license: other
tags:
  - image-classification
  - ai-generated-image-detection
  - lorc
  - dinov3
pipeline_tag: image-classification
---

# mLoRC (Modulated-LoRC)

An [LoRC](https://arxiv.org/abs/2608.20882) (Low-Rank Collapse) AI-generated-image
detector: a frozen **DINOv3 ViT-H+/16** backbone + LoRA adapters, an orthogonal
decomposition of the patch tokens against the [CLS] token, and a Low-Rank
Attention Block on the residual subspace, fine-tuned on the full
[DDA-Training-Set](https://arxiv.org/abs/2608.20882) (118,287 real/fake
pairs) with **pair-aware energy augmentation** — a training-time trick that
randomly rescales each pair's residual-subspace magnitude (simulating
different image compositions/energy bands) while mathematically guaranteeing
the real>fake energy ordering *within* every pair is preserved exactly.

`attn_rank=64`, LoRA `rank=32/α=32` (folded into the backbone weights below,
not shipped as a separate adapter).

## Why pair-aware, not per-sample

An earlier version of this augmentation drew an independent random scale for
every image, real and fake alike. That let a real image get scaled down
while its own paired fake got scaled up in the same batch — a real,
quantified risk (23.75% instantaneous real/fake energy-inversion rate per
augmented draw, vs. a 3.80% natural baseline). This checkpoint's training
draws **one shared scale factor per real/fake pair** instead — proven, not
just observed, to leave the inversion rate exactly at the 3.80% baseline,
since scaling both sides of a ratio by the same factor can't flip its sign.

## Results (full 30,000-image WildFake eval)

| | Clean BAcc | Clean AUC | Full (transformed) BAcc | Full AUC |
|---|---|---|---|---|
| v2 baseline (no aug) | 95.01% | 0.9911 | 91.92% | 0.9739 |
| **mLoRC** | **96.57%** | **0.9929** | **92.65%** | 0.9723 |

Biggest gains: real-photo groups that were previously the model's weakest
point — celebahq (85.4%→94.7% under transforms), ffhq (84.3%→94.4%). Full
per-generator breakdown, throughput benchmarks, and training details: see
[Buxt-Codes/AIGI-mLoRC](https://github.com/Buxt-Codes/AIGI-mLoRC) — the code
that loads this checkpoint lives there, not in this HF repo.

Known regression: Imagen (Google) under transforms, 91.9%→90.9% — the one
generator where this trick's real-photo gains don't fully offset a drop in
raw recall on that specific generator (93.7%→86.1%).

## Usage

Clone [Buxt-Codes/AIGI-mLoRC](https://github.com/Buxt-Codes/AIGI-mLoRC), set
up its `requirements.txt`, then:

```python
from mlorc import ModulatedLoRC

model = ModulatedLoRC.from_pretrained()      # pulls mlorc-full.pt from this repo
result = model.predict_image("photo.jpg")
print(result)   # {"label": "fake", "p_fake": 0.93, "p_real": 0.07}
```

or, for a whole directory of images:

```bash
python predict.py --input_dir <path/to/images> --output results.json
```

This repo is **private** — pass a token (`from_pretrained(hf_token="hf_...")`
or set `HF_TOKEN`/run `huggingface-cli login`) to access it. Getting access
here is the *only* gate: unlike the base DINOv3 checkpoint
(`facebook/dinov3-vith16plus-pretrain-lvd1689m`, gated with **manual review**
by Meta), loading this model never touches that repo at all — see `Files`.

## Files

- `mlorc-full.pt` (~1.6GB, bf16) — **one self-contained checkpoint**: the
  DINOv3 ViT-H+/16 backbone with the trained LoRA adapters already folded
  into its weights (`peft`'s `merge_and_unload()`, verified numerically
  exact before shipping — max output difference vs. the pre-merge model was
  1e-12, pure floating-point noise), plus the Low-Rank Attention Block and
  classifier head.
- `dinov3_config.json` — DINOv3's architecture metadata (hidden size, layer
  count, etc.) mirrored here, **not its weights**. Together with
  `mlorc-full.pt` this means loading the model never needs access to
  `facebook/dinov3-vith16plus-pretrain-lvd1689m` at all — verified with that
  repo's local cache deliberately removed before loading, output still
  bit-identical. No adapter file, no second HF repo, no Meta gate to clear.