Origin Bench
How much of this was made by a machine?
Three tests. An image. A video or a voice clip. A piece of text, split into what was written by a model and what was edited by one. File marks are read here. A model reading runs only when you ask for it, and it is an estimate. You can also load a public webpage and test its readable text.
Image test
Drop an image
The original file is scanned for generator and camera marks. A smaller copy is sent only if you ask for a model reading.
No score yet
—
Drop an image. A generator mark is scored at once. The model reading is separate.
What the percentages mean
A generator name, a diffusion parameter block, or a generative content credential is treated as file evidence. A camera or phone identity with no generator mark pulls the other way. Neither one sees pixels.
The model reading looks at the picture, the sampled frames, or the words. It can be wrong. When it and a hard file mark disagree, the file mark is the firmer of the two.
Voice timbre is not certified. If the soundtrack can be sampled, the voice figure is about the spoken words. It becomes a file figure only when the container names a voice or avatar engine.
The before-and-after text test is counted from the difference between the two drafts. It is not guessed. The single-passage test is an estimate of two separate axes: composed by a model, and later edited by one. A loaded webpage is reduced to its readable text. Menus and scripts are left out.