Elo ranking and related infrastructure

[curiousjp]

The current star based ranking system is good, but I think a more precise system that focuses on ranking images competitively against one another would be helpful for some use cases, especially culling folders to certain numbers of images.

I suggest: two new fields, Elo and the datestamp of the last Elo change event. Both are sortable in the main view and can be used to filter. A new context item to reset a record to the default Elo. A new two-up view for images that has hotkeys for awarding an Elo win or draw between the images. A way to launch this view using the currently filtered or selected images, shuffled.

Thoughts? Happy to try and assist on building it, but thought I should run it by you first.

[tachyon-beep]

Even more so if you could rank things multidimensionally. Kind of like a visual MDCA.

[RupertAvery]

Not familiar with Elo, but just to try to get an idea:

So you need to show two images, and the viewer decides which one wins by just selecting one. The winner takes points from the loser. Then, the next image takes the place of the loser?

I've also thought about multiple rankings, but kind of like attribute based, realism, prompt adherence, lighting, probably user-defined attributes with some base attributes, but having so many things to tweak would be time consuming for the user.

[curiousjp]

So you need to show two images, and the viewer decides which one wins by just selecting one. The winner takes points from the loser. Then, the next image takes the place of the loser?

You’ve got the basic idea. Using the existing scores for each competitor the algorithm attempts to predict who should win based on historical performance and the likelihood of this happening. The amount of points transferred depend on this - if the strong favourite is chosen, few points move from the loser to the winner. But a big upset moves a lot of points - this can help a strong newcomer find their level more quickly without having to beat half the field.

In terms of selecting the next matchup, there is no prescriptive rule. You could look for the scores that are closest to each other (likely to yield the most information?), that have had the fewest ranking events (needs to track an extra variable) or something else.

I would be thrilled if you looked into this so please let me know if you need some help. I would love to just offer up a pull but I know essentially nothing about GUI programming :/