GaussMass
Reads the statistical fingerprints a generator leaves in an image: sensor-noise residue, frequency artefacts and compression history.
- 93.89 %
- recall on edited synthetic images
Images, video, live calls and audio. Four models built in-house, one for each kind of media, combined into a single verdict with the evidence behind it. Each model is measured on named slices and re-scored whenever a new generator family appears.
Reads the statistical fingerprints a generator leaves in an image: sensor-noise residue, frequency artefacts and compression history.
Scores video frame by frame and across time, catching what a single frame hides.
Checks a live camera session while it is happening, for a real and present person on the other side.
Returns a calibrated likelihood that speech was generated, and refuses audio it cannot judge.
A single pooled accuracy hides where a detector fails. Each figure below comes from a named slice, and train and validation sets draw on different generator families. How the numbers are made →
Detectors fail most often when a new kind of generator appears, so each family is measured as its own slice and never pooled into one accuracy. Slice-level results by model version accompany a pilot. Read the white paper ↗
Periodic traces that upsampling and diffusion pipelines leave in the frequency domain.
Statistical fingerprints of a generator, read from noise residue and texture.
Generated images that were cropped, compressed or retouched afterwards.
A face placed on another body, in a still or across a clip.
Speech and mouth movement that fall out of step over time.
Light on a face that disagrees with how the subject moves.
A face whose identity shifts between frames or cuts.
A screen, a print or a mask held up to a live camera.
Synthetic video fed into a live session in place of the camera.
A calibrated likelihood that a voice was generated, not recorded.