All the experiments are available here: (50MB)
This is a good image for this experiment as it has several objects for the API to detect: person, mug, gun, table. And text for us to test how to hide. And Googles Cloud Vision can also interpret context such as memes to help gauge the safe search rating.
I tried using a glitch feature first. It still got everything from the face, not that that is relevant in the image, and worked to make the results only slightly less accurate. Googles system is proving far more sophisticated than thats being used on WeChat and Weibo.
Tried overlaying the sex image from Tom White, proved effective at masking the original image but brought in the Racy aspect from the sex image. By itself this image is 100% Racy.
Next I tried using the melt feature from Befunky.
And after a couple of attempts it worked. Managed to obscure a gun. But this technique is quite temperamental with the specific amount of the effect applied and the results. Also not ideal for a lot of images. Overlaying other images simply resulted in a lot of package goods being identified.
[Image source](https://twitter.com/dear52stranger/status/1197908160868077573]) Tweet archive
You can see here how sophisticated it is with relating the image to content online.