![ｲﾍﾞﾝﾄを必死に進めてるRuan[ルアン]](https://pbs.twimg.com/profile_images/1101108091423092737/sWvbm19H_normal.jpg)
More samples from today: https://imgur.com/a/zSzwjAY IRC reaction: 'holy shit that is insane. seriously this is fucking insane, when did it get so good' Me: 'there is no fire alarm' ( ‾ ʖ̫ ‾)
Show this thread
MFW when I look up from my GAN at the latest psych paper but then realize it's a Registered Report.pic.twitter.com/7jJw17Ekqk
-
-
-
Another interesting thing: the figure 8 psi interpolation. 0/middle=global mean - a brown-eyed brown-haired girl. Fair enough. 1 vs -1 are opposites. What's the opposite of gray eyes/short brown hair? red eyes/long purple hair, apparently. red eyes/long red? yellow/short blue.pic.twitter.com/9HStf2Jmoy
Show this thread -
Expanding psi to make the transitions clearer: psi=[1, 0.7, 0.5, 0.4, 0.2, 0.1, 0, -0.1, -0.2, -0.4, -0.5, -0.7, -1]. (Head pose reversal/negation is particularly obvious with this version.)pic.twitter.com/8ymUpp8DW9
Show this thread
End of conversation
New conversation -
-
-
Current StyleGAN model if anyone wants to use a good-quality but unconverged anime-face StyleGAN:https://mega.nz/#!YUskBQJa!Zi6yR-RF5-UmVs82LvanN06D_cI4Z5JGY1eVzem6gZE …
Show this threadThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I'm really enjoying these finetuning runs: taking the full face-corpus StyleGAN model and retraining it on just a specific character's corpus. Here's 2.5h (~5 epoches) of retraining on just Holo faces.pic.twitter.com/ubuVDgWd7D
Show this thread -
If 1-2h is enough to finetune on a character, what happens if you run for 7+ hours? Overfitting? No - it gets better *and* worse: eg it's picked up more global structure like eating apples (!) but at the cost of horrible noise artifacts (???). I have no explanation for this.pic.twitter.com/L5MtxQqcRV
Show this thread -
~100 of my best, hand-selected Holo StyleGAN samples: https://imgur.com/a/xXuKwYV Also, ~130 of my best Asuka Evangelion samples:https://imgur.com/a/7R8IRVh
Show this thread -
Someone asked for bizarre & bad samples, so here's ~1000 Asukas with the truncation cranked up to psi=1.2: https://mega.nz/#!gEFVwAoK!qYrejFI0w1g1BzeuI-st5ajaQLoVFOZsj5j_OTREp1c … Enjoy, I guess?
Show this thread -
I say they're bad but many are legitimately cool or interesting. Setting psi=1.2 just makes the quality range huge... eg these 4 are striking. (BTW
@EvaMonkey have you been looking at any of these Asuka samples? So lulzy.)pic.twitter.com/BH8MPe9Sn7
Show this thread -
More 'Bad Ganime'! Psi=1.2 on the latest face model, 1000 samples + interpolation videos: https://mega.nz/#!2DRDQIjJ!JKQ_DhEXCzeYJXjliUSWRvE-_rfrvWv_cq3pgRuFadw …pic.twitter.com/Nj5Jj65m2v
Show this thread
End of conversation
New conversation -
-
-
The StyleGAN anime face interpolations are solid: - https://www.dropbox.com/s/sqt761i9o1osjxq/2019-02-11-stylegan-danbooru2017faces-interpolations-fine_503.mp4?dl=0 … - https://www.dropbox.com/s/z0iqlkj4e0nxkfc/2019-02-11-stylegan-danbooru2017faces-interpolations-coarse.mp4?dl=0 … - https://www.dropbox.com/s/ux9p9wtu88v8c4t/2019-02-11-stylegan-danbooru2017faces-interpolations-random_grid_404.mp4?dl=0 …pic.twitter.com/7T7hoCDXMg
Show this thread -
Alright ya filthy animals, have some more videos (Holo/_Spice & Wolf_): - https://www.dropbox.com/s/qgk9pxr1n6hrlzz/2019-02-11-stylegan-holo-011370-interpolate.mp4?dl=0 … - https://www.dropbox.com/s/14lt9ccg891twvo/2019-02-11-stylegan-holo-011370-random_grid_404.mp4?dl=0 … - https://www.dropbox.com/s/roywal7e23c5iwf/2019-02-11-stylegan-holo-011370-fine_503.mp4?dl=0 …pic.twitter.com/ooPiXTAxrQ
Show this thread -
Hi everyone thanks for reading my tweet! Be sure to check out my Soundcloud and buy the t-shirt:https://www.patreon.com/gwern
Show this thread
End of conversation
New conversation -
-
-
I have a confession. When I submitted my best Asuka & Holo samples to Reddit (https://imgur.com/a/7R8IRVh https://imgur.com/a/xXuKwYV ), I may not have been entirely candid about their nature: - https://www.reddit.com/r/evangelion/comments/apmkjm/brighten_your_monday_with_some_asukas_album_of_130/ … - https://www.reddit.com/r/SpiceandWolf/comments/apazs0/my_holo_face_collection/ … https://www.reddit.com/r/SpiceandWolf/comments/apbz6r/all_those_cropped_holo_faces_uprimarypizza_posted/ … ᕙ( ͡° ͜ʖ ͡°)ᕗ
Show this threadThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
An experimental run of StyleGAN on the full Danbooru2018 SFW 512px image corpus is pretty trippy-looking so far. Wonder if the initialization from the faces StyleGAN is why it looks so Boschian...pic.twitter.com/o0BNTceakz
Show this threadThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
People keep asking about reversing StyleGAN to control/generate variants on an existing natural image. This right now seems to be the best thing available:https://www.reddit.com/r/MachineLearning/comments/aq6jxf/p_stylegan_encoder_from_real_images_to_latent/ …
Show this threadThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
If you don't want to browse through the dumps, you can see random face examples at http://www.thiswaifudoesnotexist.net/index.html https://twitter.com/gwern/status/1097973862392844290 …
Show this threadThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Looks like someone's hooked up StyleGAN to latent-variable classifiers, allowing interactive control of the visual attributes of each face: https://www.reddit.com/r/AnimeResearch/comments/asl4gw/fully_custom_waifu/ … OP's not answering any questions, though. Anyone know where this comes from? Not finding anything in Google.
Show this threadThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
these are really good images
-
Aren't they? I literally burst out laughing when I finally got `generate_figures.py` working and could look at the interpolations.
-
-
Some of them I look at and think wow this looks like an image of a high quality anime. Then I look closer at the image and the shoulders are messed up.
-
Yes, both ProGAN and StyleGAN struggle with the edges and background because they're obscured and highly-varied. The faces are always in clear focus in the foreground, ofc. I imagine shoulders should get better toward the end - not even at 512px yet!
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.