sadracoding.ir GitHub Hugging Face
A RESEARCH DEMO

Real face,
or a machine’s?

Upload a face image. The model weighs visual detail and hidden frequency patterns, then gives its best experimental guess.

Try the demo How it works
Clean accuracy
93.27%
Test images
10,905
Model input (px)
224
Swin FFT · log|F(u,v)| Fusion
01TRY THE MODEL

Give it any face image.

RUNS ON DEMO SERVER

Drop a face image here, or click to browse

PNG, JPG, WEBP · max 12 MB

Spectrum view is computed in your browser.

Your image is only used for this one request — nothing is saved on disk.

ANALYSIS RESULT

Once you pick an image, the result shows up right here.

Go ahead, choose a face image to get started.

Heads up — this is just a model prediction, not definitive proof.

Best with face images · Experimental model

02HOW IT WORKS

Three steps, two kinds of evidence.

The model examines details and frequency patterns separately, then combines the clues.

01SWIN BACKBONE

Look at image details

The Swin model examines shapes, textures and patterns at different scales.

02FFT · LOG-MAGNITUDE

Look for fine patterns

A Fourier transform reveals frequency patterns that are hard to see by eye.

03BINARY CLASSIFICATION

Bring the clues together

Both views are fused to decide: “human” or “artificial”?

03REPORTED EVALUATION

How it performed in testing.

View raw metrics↗
ACCURACY ON CLEAN IMAGES
93.27%
Confusion matrix
Pred. artificialPred. human True artificial
5,355
137
True human
597
4,816

From one recorded evaluation run

IMAGES TESTED
10,905
Artificial 5,492Human 5,413

Face images in two categories

ADAPTIVE FAST PATH
99.06%

Images that finished early via the fast path

Full path6.60 ms
Adaptive6.24 ms
ACCURACY UNDER IMAGE CHANGESaxis starts at 88%
  • Clean
    93.27%
  • Gaussian blur 5×5
    92.86%
  • Half resolution
    92.11%
  • JPEG quality 50
    91.30%
  • Gaussian noise σ=15
    91.24%

Synthetic degradations on the same test set

04PROJECT LIBRARY

Explore the project.

Browse the code, read the report, or explore the dataset.

sdxldd.irOpen the live demo GitHub · SDXL-Deepfake-DetectorSource code and project files Hugging Face · SDXL-Deepfake-DetectorModel checkpoint and files