search

keptan / superCuteGrab

Booru image tag-fetching and ranking with Microsoft TRUESKILL and ELO
11 stars 0 forks source link

Applied usage in Neural Image Rankers #2

Open DonaldTsang opened 4 years ago

DonaldTsang commented 4 years ago

Since https://github.com/KichangKim/DeepDanbooru exists, and there are other networks like:

Is it possible to try and pinpoint the application of SuperCuteGrab (SCG)?

Also another question: how hard would it be to create a decentralized and gamified ranking system, between multiple co-training neural networks, and updating human ground truth data?

keptan commented 4 years ago

Combining tag-estimation data with the trueskill data would be interesting. Assuming the tag skills were predictive it would allow to predict an images 'score' before it was ever rated using humans.

I think the core problem with this is the effectiveness of the tags at actually predicting an images quality. For the most part the tag data is very low quality because of the low information density

When you compare two images, each with one hundred tags you barely get any entropy increase, it would take literally millions of rankings before you got close to an accurate picture. Tag data is very noisy.

I think artist metadata is more useful. However also much harder to be estimated by a neural network....

As for the 'decentralized' idea, that is something I was starting here: https://github.com/keptan/cureMaid but didn't put much work into yet

DonaldTsang commented 4 years ago

I think the core problem with this is the effectiveness of the tags at actually predicting an images quality. For the most part the tag data is very low quality because of the low information density When you compare two images, each with one hundred tags you barely get any entropy increase, it would take literally millions of rankings before you got close to an accurate picture. Tag data is very noisy.

There is a team of programmers at https://discord.gg/vy8CUB4 regarding a decentralized AI tagging system (made by "that guy") where people are encouraged to manually tag images to re-train AIs regularly. Do you have Discord, since these projects may have collaborative potential?

I think artist metadata is more useful. However also much harder to be estimated by a neural network....

Unless AI can handle art style and textures, that would be hard.

As for the 'decentralized' idea, that is something I was starting here: https://github.com/keptan/cureMaid but didn't put much work into yet

Good to know that we are on the same page.