Fake-Inversion:

Learning to Detect Images from Unseen
Models by Inverting Stable Diffusion

CVPR 2024

Massachusetts Institute of Technology, Google Research

A New Gen-AI Detector


Prior fake image detection methods [1, 2] learn to detect fake images by simply training a classifier on the Original images. We wish to leverage a pre-trained diffusion model (Stable Diffusion 1.5) to train a better detection model.

To do this, we obtain additional input features by using DDIM Inversion and then use the Original, Inverted, Reconstructed together as the input to our model.

Despite only using features extracted from a weak generator (SD v1.5), our method generalizes remarkably well and can detect fake images from modern propriatary (DALL-E 3, Midjourney, Imagen, etc.) and open-source (Kandinsky, PixArt-α, etc.) generators that were never seen during training.

A New Evaluation Benchmark

LAION
“Paintings By 刘诚 Liu Cheng Fine Art and You Painting”
DALLE‑3
Midjourney v6
PixArt‑α
Playground 2.5

Existing evaluation methods for GenAI detectors only use LAION images as the "real" samples in the test set, despite generated "fake" images typically having very different "styles" than the "real" LAION images (see above). Unfortunately, this often results in a style mismatch between the "real" and "fake" test images, calling into question whether or not the model is actually detecting "realness" or merely some kind of style bias.

To help alleviate these concerns, we introduce a new set of benchmarks intended to show how well a model would perform in-the-wild. We do this by searching the (Pre-GenAI) Internet for images that most closely match the style and content of the "fake" images obtained from the generative model in question.

When collected this way, the distribution of "real" test images is much closer than LAION to that of the "fake" images from the generator.

Image Browsing


Please use the following links to view previews of our new test sets. Fake images are on the left, and the corresponding Real images found via reverse image search are on the right.

DALL-E 3

Midjourney

Kandinsky 2

Kandinsky 3

PixArt-α

Playground 2.5

SDXL

SDXL DPO

Seg-MoE

SSD-1B

Stable-Cascade

Segmind Vega

Wurstchen 2

BibTeX

@article{cazenavette2024fake,
  author    = {Cazenavette, George and Sud, Avneesh and Leung, Thomas and Usman, Ben},
  title     = {{Fake-Inversion}: Learning to Detect Images from Unseen Models by Inverting Stable Diffusion},
  journal   = {CVPR},
  year      = {2024},
}

References

[1] Wang et al., CNN-generated images are surprisingly easy to spot...for now (CVPR 2022)
[2] Ojha et al., Towards Universal Fake Image Detectors that Generalize Across Generative Models (CVPR 2023)