Funny related story! At Reddit we built a tool to create thumbnails for webpages. The algo was basically find the biggest “squareish” image and then remove a row or column of pixels. The row/col was chosen based on which one left the image with the most entropy. We kept doing this until the image was a square.

That algo basically cuts off women’s heads and leaves just their cleavage.

Obviously we had to adjust it to try and avoid that, but even today you’ll see thumbnails on Reddit that are just cleavage.

I thought that a face would contain more entropy than two flesh mounds! Color me surprised. Could the algorithm be tweaked to favor removable from the bottom part of the image, since humans generally perceive from top to bottom? Most images contain the 'meat' at the top or middle, I'd assume?

We thought so to! But it turns out that a nose, mouth, neck, and chest have more entropy than a face.

I don’t remember exactly how we fixed it but I think biasing towards the top was involved.

Link from yesterday shows some good examples of the "algo" (seam carving):

https://avikdas.com/2019/07/29/improved-seam-carving-with-fo...

Seam carving gone wrong?

Of course, along with the very-fucking-annoying Youtube Face [0]. It's so ordinary to mess with human psychology, anything to increase click through rates!

[0] https://boingboing.net/2018/04/11/youtube-face-the-same-thum...

It seems like most of the examples of that are "reaction" videos, and if you're trying to sell a reaction video, you show the reaction, right?

I see it everywhere, not just on reaction videos.

I can count the number of reaction videos I've seen on one hand but the YouTube face is all over my homepage.

The next iteration to this is out, it's using extreme Snapchat-esque filters to comically exaggerate the size of the eyes and mouth of a reaction face.

It's mesmerisingly disgusting, I wonder what the next steps are. More digital augmentation awaits?

> I see it everywhere, not just on reaction videos.

I have seen them in in-store advertising in supermarkets.

> I wonder what the next steps are. More digital augmentation awaits?

Virtual/retouched streaming personas. It's already a thing in japan and among furries.

What software is being used to create virtual personas? Am I right in understanding that this work isn't being done by Western companies?

I haven't dug deeply, but there seems to be a whole ecosystem. An open source avatar data format[0], full body motion capture software[1], this stuff integrates with standard 3d software environments such as unity/unreal/blender/maya/max. So does commercial software[2]. For the hardware it can be a simple web/phone camera on the low end (giving you a talking head) over camera + VR trackers + an iphone for the face sensor to high end rigs with multiple cameras and mocap suits.

[0] https://vrm.dev/en/ [1] https://github.com/sh-akira/VirtualMotionCapture [2] https://www.youtube.com/watch?v=hJ_5d_1NHCQ

I don't think it's 100% the same thing (VRM appears to be full body, not just face), but FaceRig[1] is a similar concept made by a Romanian studio.

[1] https://facerig.com/