abacusai/Smaug-AI-Image-Detector-GGUF

---

1. About

The official GGUF build of Smaug-AI-Image-Detector for llama.cpp, converted by Abacus.AI from the released weights and checked image by image against the PyTorch model.

The detector is an image encoder followed by a small classification head. llama.cpp runs the encoder and returns an image embedding; head.json (768 weights and a bias) and smaug_image_detect_gguf.py turn that embedding into the probability that the image is AI-generated, using the same 528 × 528 crops as the PyTorch release. A GGUF file on its own returns an embedding, not a verdict.

2. Files

File	Size	Use
Smaug-AI-Image-Detector-f16.gguf	558 MB	Language model (f16): reference, closest to PyTorch
Smaug-AI-Image-Detector-Q8_0.gguf	310 MB	Language model (Q8_0): recommended
mmproj-Smaug-AI-Image-Detector-f16.gguf	367 MB	Vision encoder (f16): reference, closest to PyTorch
mmproj-Smaug-AI-Image-Detector-Q8_0.gguf	226 MB	Vision encoder (Q8_0): recommended
head.json	10 KB	Classification head: 768 weights and a bias
smaug_image_detect_gguf.py	4 KB	Client: crops, llama.cpp embedding, head

Use a language model and a vision encoder (mmproj) of the same precision: f16 with f16, or Q8_0 with Q8_0.

3. Accuracy vs. the PyTorch Release

Language model + vision encoder	Download	AUC	Average probability difference	Largest probability difference	Verdicts changed	Seconds per image (CPU)
f16 + f16	925 MB	0.9722	0.0079	0.1714	0 of 200	2.0
Q8_0 + Q8_0 (recommended)	536 MB	0.9720	0.0065	0.1981	0 of 200	2.4

Measured on 200 real-world images from the OpenFake Reddit test split (100 real photos, 100 AI-generated), three crops each, against the PyTorch release on the same images (AUC 0.9723). "Verdicts changed" counts images that land on the other side of the 0.5 threshold. The token sequence and pixels are identical to the PyTorch release; the small remaining differences come from floating-point arithmetic and are largest for images close to the threshold. CPU timings: 16-core (32-thread) Intel Xeon, llama.cpp default threads, f174553. Full benchmark results are on the main model card.

4. Usage

# 1. Download one language model, one vision encoder, the head and the client
hf download abacusai/Smaug-AI-Image-Detector-GGUF \
  Smaug-AI-Image-Detector-Q8_0.gguf mmproj-Smaug-AI-Image-Detector-Q8_0.gguf head.json smaug_image_detect_gguf.py \
  --local-dir smaug

# 2. Serve the image encoder with llama.cpp
llama-server -m smaug/Smaug-AI-Image-Detector-Q8_0.gguf --mmproj smaug/mmproj-Smaug-AI-Image-Detector-Q8_0.gguf \
  --embeddings --pooling mean --image-max-tokens 140 --port 8080

# 3. Score images
pip install pillow requests
python smaug/smaug_image_detect_gguf.py photo.jpg another.png

from smaug_image_detect_gguf import SmaugImageDetectGGUF

detector = SmaugImageDetectGGUF("http://127.0.0.1:8080")   # the llama-server from step 2
detector("photo.jpg")                                       # -> float in [0, 1]
detector(["a.jpg", "b.png", pil_image])                    # -> list of floats
detector("photo.jpg", crops=1)                              # 3x faster

--pooling mean and --image-max-tokens 140 are required: they match how the detector was trained. 0.5 is a good default threshold; raise it when a false alarm is costly.

5. How It Works

1. Crops	Shorter side resized to 576 px, then three 528 × 528 crops along the long side (client)
2. Embedding	EmbeddingGemma-2 vision encoder and language model, mean-pooled and L2-normalized, 768 dimensions (llama.cpp)
3. Score	Linear head on each crop's embedding, logits averaged, then a sigmoid (client, head.json)

6. License

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

7. Citation

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