| [ | Date | | | 2022-11-29 23:31 -0500 | ] |
In 2022, fun or useful picture-related artificial-intelligence-based Web applications became available. One of them is Stable Diffusion, which consumes textual descriptions and produces matching pictures.
This post is not about Stable Diffusion. It’s about Cleanup.pictures, which lets its user remove unwanted objects from pictures by applying a brush over those objects. The system takes care of filling in the void.
I thought it would be interesting to see what object removal would do to a very iconic picture: that of Stalin next to Nikolai Yezhov, a version of which exists that has been retouched to remove Yezhov, very non-subtle way. In my opinion, the framing is a strong hint that the picture was altered, as it gives lot of room empty of the characters meant to be depicted, a choice that seems unusual to me for such formal pictures.
To my surprise, the higher-resolution input I thought would be best doesn’t get processed well; the ramp behind Yezhov is reconstructed so poorly that it’s very obvious that something happened to the picture.
The object removal software does a reasonably good job of removing Yezhov on the alternate input, which has wider framing, lower resolution and higher contrast.
I wonder if the fact that the original picture and its manually altered version are famous historical artifacts mean that the artificial intelligence model had access to it.
While we’re there, it’s not very time-consuming to try other alterations to the original picture.
The void left by Stalin becomes a blur that merges with part of the remaining person on his right.
Quick links:
David Haguenauer
<www@kurokatta.org>
$Id: yezhov-cleanup.malblog,v 1.2 2022/11/30 04:31:25 homer Exp $