[ home / board list / faq / random / create / bans / search / manage / irc ] [ ]

/hydrus/ - Hydrus Network

Bug reports, feature requests, and other discussion for the hydrus network.

Catalog

Name
Email
Subject
Comment *
File
* = required field[▶ Show post options & limits]
Confused? See the FAQ.
Embed
(replaces files and can be used instead)
Options
Password (For file and post deletion.)

Allowed file types:jpg, jpeg, gif, png, webm, mp4, swf, pdf
Max filesize is 8 MB.
Max image dimensions are 10000 x 10000.
You may upload 5 per post.


New user? Start here ---> http://hydrusnetwork.github.io/hydrus/

Currently prioritising: simple IPFS plugin


File: 1449571303650.png (104.98 KB, 1366x768, 683:384, neat.png)

 No.1553

Reposting from >>>/tech/457184

Maybe relevant to your interest.

Japan did it.

Automatic anime images tagging with neural networks.

THE FUTURE IS NOW

http://illustration2vec.net/

http://illustration2vec.net/papers/illustration2vec-main.pdf

 No.1554

File: 1449598315609.jpg (39.84 KB, 478x640, 239:320, 872c7f2c44e2085dbfd6e3060c….jpg)

Thanks for this link, it looks great! You might have seen that it can also run offline (which is how I prefer this sort of thing) and has a python module, so it would be very easy for me to integrate. I plan to add a 'suggested tags' control to pop up aside the manage tags dialog, which this 'tag feed' could optionally plug into very easily.

Furthermore, it might be something we can feed back into in future, improving their pre-trained models with our own growing collection of tag-feature metadata. I am not sure what data they used to generate their models–I assume a sample of some booru–but they say it knows about 1,539 different tags, as listed at http://illustration2vec.net/models/tag_list.json.gz (including "no bra" and "cum in pussy", jej). If this is just a proof of concept, and there is some easy way to train your own models, I think we can extend it to do a bit more than that!

>tfw within ten years, advanced neural networks will spend trillions of CPU cycles hyper-accurately scanning all human media for different types of semen-related metadata.

I will add this to the list of new things to add, and put it up on the vote list once I have time to work on something new.


 No.1555

Fucking awesome.

>anime images

It's a shame that it won't run on my unorganized to fuck reaction folder, but it's still useful for mega:///#F!Y0pTnZiY!g0ALqjpneHo-Lr2cJJaBuA


 No.1556

File: 1449611558311.jpg (38 KB, 294x333, 98:111, 1444276459170.jpg)

>>1554

Hydrus dev you really outdo yourself. I hope you know it's appreciated. I was just sharing it out of interest but you considered adding it as a feature and even contributing back, that's really nice.

Also I knew that not organizing my image collection manually would pay off eventually. h-heh


 No.1561

File: 1449706926655.jpg (74 KB, 529x386, 529:386, 667cb9f288323315944a8f20fc….jpg)

>>1556

Thank you. I am glad to help out a fellow Anon.


 No.1568

File: 1449951770587.jpg (22.29 KB, 396x286, 18:13, 1419478099637.jpg)

>>1553

>>1554

Fuck yeah, I heard about this the other day and came here to mention it, and based Hydrus dev is already on the case.

I look forward to seeing it implemented in the future. 'need mo tag metadata fo mah program


 No.1572

File: 1450069742755.jpg (129.29 KB, 500x340, 25:17, 1363370424761.jpg)

>>1554

>hydrus dev already planning on implementing this


 No.1670

Would be awesome to see this implemented.

Some kind of ability to retrain with new data sets, if at all feasible, would greatly add to the value too. The current net is pretty solid for anime stuff, but if you're trying to work with a different kind of image set it's only so so. And even with anime, the data used to train the current system was presumably a lot smaller than it could have been, due to being a proof of concept and all.




[Return][Go to top][Catalog][Post a Reply]
Delete Post [ ]
[]
[ home / board list / faq / random / create / bans / search / manage / irc ] [ ]