Original · Standard map
Original image
The original scene has almost no Standard-map response; the bear that will be changed is unmarked.
HANDS-ON GUIDE
Reading suspect-region maps
In this tutorial we demonstrate the different ways that both positive and negative examples of shallowfakes can be detected, as well as the type of complicated images that can cause false positives. You can download these examples to try them in the scanner yourself.
Because of the way some deepfake images are constructed, they may show large masked regions. That's not a bug: deepfakes can have different structure than real images, and that is also exposed by the algorithms we use for shallowfake detection.
Nine upload-ready images, three edit masks, and source notes.
01
Three edits with quiet originals
Scan each original and its edited version. The Standard map is quiet over the target in the original, then responds where the edit was made. Open the supplied edit mask after comparing the maps.
A bear altered
The original map stays quiet over the bear on the left. After a localized Photoshop edit changes that bear, the map marks the edited area.
Download image ↓
Edited · Standard map
Edited image
The response appears over the changed bear and follows most of the edited area.
Reveal the edited area
The translucent red area is the dataset's edit mask. It marks the area that was changed, independently of the scanner response.
Download the edit mask ↓An elephant altered
The left elephant is quiet in the original map. An edited version changes that elephant, and the map responds over it.
Original · Standard map
Original image
The original scene produces a quiet Standard map, including the left elephant.
Edited · Standard map
Edited image
The edited map covers most of the changed elephant and remains largely quiet elsewhere.
Reveal the edited area
The translucent red area shows the supplied edit mask for the left elephant. Compare its shape with the map above.
Download the edit mask ↓A skier's skis changed
The original snow scene is nearly quiet. In the edited version, the map follows the changed skis without broadly marking the snow.
Original · Standard map
Original image
The Standard map shows almost no response on the original skis or the snow.
Edited · Standard map
Edited image
The response traces the changed skis and leaves most of the scene unmarked.
Reveal the edited area
The translucent red area is the supplied edit mask for the skis. Compare it with the map response before opening the mask.
Download the edit mask ↓
02
False signals in control scenes
These copies received no local content edit during preparation. Scan each one and inspect the Experimental map. Its additional noise analysis is especially prone to false positives.
No local edit · Experimental map
Dense foliage
The experimental map marks leaves and branches in a control image. Texture alone can produce suspect regions.
No local edit · Experimental map
Clouds and reflections
Broad responses cover parts of the sky, shoreline, and water. A large colored area can still appear without an introduced alteration.
No local edit · Experimental map
Natural light transitions
Scattered responses occur near bright sky, sunlit cliffs, and vegetation. Compare each marked region with the scene before treating it as an edit.
03
Sources and use
The edit pairs come from a public research dataset. The control scenes have public source pages.
- Three edit pairs
- TGIF: Text-Guided Inpainting Forgery Dataset by Hannes Mareen and collaborators, distributed under CC BY-SA 4.0. The source photographs are MS COCO images, which the dataset credits under CC BY 4.0. These are Photoshop inpainting edits with supplied masks; the originals and edits retain their native PNG pixels.
- Dense foliage
- Red-shouldered Hawk by Heather Henkel / USGS. Source page states Public Domain. The tutorial uses a resized, uniformly encoded copy as a control scene.
- Clouds and reflections
- Sunset over Lake Anne by Jessica Rodysill / USGS. Source page states Public Domain.
- Natural light transitions
- Sunset clouds above Grand Canyon by Emily Palmquist / USGS. Source page states Public Domain.
The three edit pairs were selected for clear control-to-edit comparisons; they are examples, not an accuracy estimate. “No local edit” labels apply to the three control copies and do not certify the original photographs' complete editing history. Those control images were re-encoded and resized for the tutorial. Use map responses as diagnostic evidence, not authoritative answers.