Pardon me, but I can't help but compare 3dbooru to danbooru and the first thing that comes to mind is this.

Most of the images in danbooru stand on their own, but most of the images on 3dbooru are part of a set. For instance I like this Tamamo

649241

but what I really like is the set of 42 images that it comes from. If I favorite just one image from the set then it is not easy to find the rest of the set. If I favorite all the images of the set that breaks the browsing interface in favorites because for browsing I'd like to see 1 Tamamo that reflects that Tamamo + ely set, 1 Shibuya Rin that reflects a certain Shibuya Rin + Sakura Ema set, etc.

It seems to me that these could be mostly clustered by grouping together images with similar attributes (character and cosplayer seem good but it's not clear to me that those tags are tagged) A secondary approach is that images in a photoset seem to often be uploaded in sequential order so clustering over time could be useful too. For instance, ᴄᴏsᴘʟᴀʏᴇʀ × ᴄʜᴀʀᴀᴄᴛᴇʀ would neatly cut out that Tamamo but I'm sure there are cases where a cosplayer does significantly different cosplays of the same character, but if the breaks can be derived from time it won't be too hard to manually segment the image stream.

It is is complicated a little by this: some of the photosets seem to have a hierarchical structure. That is, this Shibuya × Sakura is wearing a certain dress 649734 and this one 649733 is wearing a bikini, there is 646855 where she is wearing a school uniform. These are all related, she is wearing the same hair ornament in the first two photos, but that big photoset probably should be split three ways. Similarly when I look through certain tags I might find 20 images from photoset A, then 35 images from photoset B, and then another 15 images from photoset A. Maybe the A should be split into A1 and A2 because often these different in setting or theme instead of clothes, for instance there is a gap between 647561 and 648532 where the same ᴄᴏsᴘʟᴀʏᴇʀ × ᴄʜᴀʀᴀᴄᴛᴇʀ is in different places, one might even want to split out 648542.

Anyhow, I found out that tags are already tagged so tags for ᴄᴏsᴘʟᴀʏᴇʀ and ᴄʜᴀʀᴀᴄᴛᴇʀ can be identified.

I am also thinking about how to give the result back.

I noticed there are tags that are tagged for photosets and the data is valuable but it's still not good enough. For instance that Tamamo photo set is embedded in a photoset that has approximately four photo sets of the same cosplayer doing different characters from Fate. There is another "photoset" with 10,000 images that all came from the same site.

Those kind of photosets are nestable, I think.

From a UI perspective however I think it might work better to use pools for photosets: pools have a distinct visual appearance on the image page that would help them be "not just another tag". Danbooru sometimes uses pools to group together a set of images that make a set.

Off the cuff I think this could create 1,000-10,000 pools. The pools would represent one layer of hierarchy (I either bunch all the Shibuya × Sakura into one pool or I insert 3 unrelated pools).

Any thoughts or objections?