I believe it makes partial guesses based on training data.

So, if that is standard colorization for those type of buttons, it might be somewhat accurate. It is possibly all wrong, of course.

The AI isn't great for detecting some objects, though it often gets a surprising amount correct. Most pics will need tweaking of the text used for colorization. Note that many pictures are currently bad for this sort of thing. Like comics, for example.

I think it is still being worked on, so I'll revisit it later after it has been improved.

To get decent color, you really have to play around with the auto-generated text. I suggest copying each set of text you like as you play around so you can see what does / doesn't work. Would be nice to find more details on how the A.I. is colorizing things.

I found that fixing the colorization for some objects would mess up the rest of the picture.

For example, there was a BW pic of a fast train. Each time I changed the color of the train, it tinted the entire picture. Some weird colors for the train made everything else look good, and some good colors for the train messed everything else up.

If you are willing to use photoshop to slice out objects, you can get better overall results.

I played with this a short while ago. Here are some examples:

• 1920 - Aerial view of Edinburgh
• Nasa - Flight Director Gene Kranz - Colorized using AI
• 1901 Advertising for Wives

I tried others. It is rather finicky. Sadly, I didn't record the text I used for each set.

For the "1901 Advertising for Wives" I used the auto settings but changed the text part. I removed the auto-added word "outhouse" and replaced it with "log cabin" to have the colorization a bit more accurate.

edit: Here is what the colorization text I used for "1920 - Aerial view of Edinburgh".

>A stock photo of wing, plane, cloud, castle, hill, grass, and forest. Small contrasting details in color. In color film.