abacusai/Smaug-AI-Image-Detector

---

1. Model Introduction

Smaug-AI-Image-Detector sets a new state of the art for open-source AI-generated image detection. It is a fine-tune of Google's EmbeddingGemma-2 by Abacus.AI and the strongest open-source detector on real-world images.

A dedicated vision model, 41% smaller than its base. We streamlined EmbeddingGemma-2 into a vision-only architecture by removing its 305M-parameter audio tower: 439M parameters instead of 744M, with no loss in accuracy.

import sys
from huggingface_hub import snapshot_download
sys.path.insert(0, snapshot_download("abacusai/Smaug-AI-Image-Detector"))
from smaug_image_detect import SmaugImageDetect     # ships with this repo

detector = SmaugImageDetect.from_pretrained("abacusai/Smaug-AI-Image-Detector")
detector("photo.jpg")                                 # 0.97 -> probability the image is AI-generated

2. Highlights

• #1 on real-world images: 0.973 AUC on OpenFake Reddit; the next best open-source detector scores 0.947.
• Catches more at the same false-alarm rate: at 5% false alarms it catches 86.7% of AI images on Reddit; the next best catches 77.4%.
• #1 on OpenFake Core (0.991 AUC) and #1 on average across all three benchmarks (0.982 AUC).
• Precise: flags only 3.0% of real Reddit photos, 1.4% of real OpenFake photos and 0.4% of real CommunityForensics photos as AI.
• Current generators: GPT-Image 2, Nano Banana Pro, Midjourney 7, Seedream 5, FLUX.2, Z-Image, Recraft, Ideogram, ERNIE-Image, Sora 2 and Veo 3 frames, DALL·E 3, Firefly 3 and more.
• Runs anywhere: at 439M parameters it runs on a single consumer GPU (about 1 GB of memory), on Apple Silicon Macs via Metal, or on a plain CPU, so images can be checked locally without sending them anywhere.

3. Model Summary

Architecture	EmbeddingGemma-2 image encoder + linear classifier
Total Parameters	439M (base model: 744M)
Audio Tower	Removed: vision-only architecture, −305M parameters
Input	Any image; scored as three 528 × 528 crops at native resolution
Output	Probability that the image is AI-generated
Hardware	NVIDIA GPU (~1 GB), Apple Silicon (Metal), or CPU
Precision	bfloat16 weights
Base Model	google/embeddinggemma-2
Adaptation	Full fine-tune

4. Evaluation

	Smaug-AI-Image-Detector	OpenFake-SwinV2	Bombek1-SigLIP-DINOv2	Thermostatic-Frontier	CommFor-384	haywoodsloan-SwinV2
AUC ↑
OpenFake Reddit	0.973	0.947	0.842	0.894	0.744	0.897
OpenFake Core	0.991	0.989	0.959	0.977	0.921	0.953
CommunityForensics	0.981	0.951	0.990	0.979	0.968	0.901
Average	0.982	0.962	0.930	0.950	0.878	0.917
AI images caught at 5% false alarms ↑
OpenFake Reddit	86.7%	77.4%	54.0%	60.8%	34.8%	59.4%
OpenFake Core	95.8%	95.0%	83.9%	87.9%	68.2%	83.1%
CommunityForensics	92.2%	81.8%	96.2%	91.6%	83.2%	74.5%
Real photos flagged as AI ↓
OpenFake Reddit	3.0%	2.9%	34.7%	16.4%	2.0%	9.5%
OpenFake Core	1.4%	2.1%	18.4%	6.6%	0.3%	1.1%
CommunityForensics	0.4%	3.0%	5.2%	17.7%	4.0%	10.6%

Bold = best in the row. Shading is relative within each row (darker = better). All models scored on the same 26,070 test images.

All open-source detectors tested

Model	OpenFake Reddit	OpenFake Core	CommunityForensics	Average	Real photos flagged: Reddit / Core / CF
Smaug-AI-Image-Detector	0.9730	0.9909	0.9808	0.9816	3.0% / 1.4% / 0.4%
OpenFake-SwinV2	0.9467	0.9886	0.9512	0.9622	2.9% / 2.1% / 3.0%
s-gehring/detectra-openfake	0.9359	0.9715	0.9535	0.9537	2.7% / 1.2% / 2.0%
Thermostatic-Frontier	0.8935	0.9766	0.9792	0.9498	16.4% / 6.6% / 17.7%
Bombek1-SigLIP-DINOv2	0.8418	0.9591	0.9900	0.9303	34.7% / 18.4% / 5.2%
haywoodsloan-SwinV2	0.8970	0.9528	0.9011	0.9170	9.5% / 1.1% / 10.6%
TensorFold/detectra-v1	0.7483	0.9330	0.9652	0.8822	5.3% / 0.6% / 3.5%
CommFor-384	0.7441	0.9214	0.9678	0.8778	2.0% / 0.3% / 4.0%
Ateeqq/ai-vs-human-image-detector	0.8665	0.9045	0.6939	0.8217	27.2% / 14.3% / 40.1%
Organika/sdxl-detector	0.6625	0.6673	0.8243	0.7180	43.9% / 22.4% / 24.8%
umm-maybe/AI-image-detector	0.7364	0.6007	0.6042	0.6471	13.8% / 17.3% / 15.3%

5. Training Approach

• Full fine-tune of EmbeddingGemma-2's image encoder with a linear classifier on its embedding.
• Native-resolution crops keep the pixel-level traces that generators leave, instead of shrinking the image.
• Robustness training with resizing, recompression and blur, so scores hold up on shared and re-saved copies.
• Hard negatives from professional photography, so a polished real photo is not mistaken for AI.

6. Limitations

• Brand-new generators, or ones built to evade detectors, may get through.
• A real photo with a small AI-edited region may be scored as real.
• A score is evidence, not proof: do not use it as the only basis for decisions about people or their work.

7. Usage

pip install "sentence-transformers[image]" pillow

from huggingface_hub import snapshot_download
import sys

path = snapshot_download("abacusai/Smaug-AI-Image-Detector")
sys.path.insert(0, path)
from smaug_image_detect import SmaugImageDetect

detector = SmaugImageDetect.from_pretrained(path)        # NVIDIA GPU, Apple Silicon (Metal) or CPU, picked automatically
detector("photo.jpg")                                     # -> float in [0, 1]
detector(["a.jpg", "b.png", pil_image], batch_size=16)   # -> list of floats
detector("photo.jpg", crops=1)                            # 3x faster

0.5 is a good default threshold; raise it when a false alarm is costly.

8. License

Apache 2.0, the same license as the base model, EmbeddingGemma-2.

9. Citation

@misc{abacusai2026smaugaiimagedetector,
  title  = {Smaug-AI-Image-Detector},
  author = {Abacus.AI},
  year   = {2026},
  url    = {https://huggingface.co/abacusai/Smaug-AI-Image-Detector}
}