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
Показване на тази нишка
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
-
-
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
Показване на тази нишка -
~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
Показване на тази нишка -
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?
Показване на тази нишка -
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
Показване на тази нишка -
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
Показване на тази нишка
Край на разговора
Нов разговор -
-
-
could create a new anime ideas for concept artist this way, also how do you set train. py to use the latest pkl file when retraining? i lost 20 hours of training due to colab deleting data after the instance is shutdown.
-
It uses the latest pkl in each run folder by default. You specify which run to resume from by editing 'resume_run_id' with the approriate subdir of results/ in 'training_loop.py'.
-
Having trouble with this as well, I get the correct "loading networks from ...pkl" message and the fakes000000.png looks correct - but then at the next tick it looks as if it had started over.
-
Are you starting at 512px (eg by setting resume_kimg=7000 in training_loop.py)? StyleGAN seems to erase all resolutions higher than the starting point for unknown reasons. Obviously, a very big problem for transfer learning.
Край на разговора
Нов разговор -
-
-
Your gan has surpassed deviantart
-
'After humanity spent thousands of years improving our anime drawings, computers tell us that humans are completely wrong... Perhaps not a single human has touched the edge of the truth of moe.'
-
Moe superstimulus accelerationism nyow
Край на разговора
Нов разговор -
-
-
How does one do this
-
Swap out the .tfrecords and keep on training.
Край на разговора
Нов разговор -
-
-
DeepSpice
-
_Spice & WAIfu_
Край на разговора
Нов разговор -
-
-
Благодаря. Twitter ще използва това, за да подобри хрониката ти. ОтмянаОтмяна
-
-
-
(5,1) is such an interesting art style. very dark and rugged.
Благодаря. Twitter ще използва това, за да подобри хрониката ти. ОтмянаОтмяна
-
-
-
Благодаря. Twitter ще използва това, за да подобри хрониката ти. ОтмянаОтмяна
-
Изглежда зареждането отнема известно време.
Twitter може да е претоварен или да изпитва моментно затруднение. Опитай отново или виж Twitter Status за повече информация.