bodys about makegirlsmoe_web

[Aixile]

Thanks for your interests.

There are two considerations.
From the model perspective, obviously, 128x128 pix is too low, I am working on the high-resolution face model now. But the training of Generative Adversiral Networks is hard and unstable, I am not sure how long will it take to finish the high-res model.

From the data perspective, a clean and well-aligned anime characters(with body) dataset is required. My current dataset is not enough for training a GAN with bodies.

I will investigate more on body generation after I finish the high-res face generation model.

[gwern]

Apropos of Danbooru2017, I've had good luck using Nagadomi's face extraction tool on it. The other day I was able to extract ~7000 pics of a single character (Asuka from Evangelion) face, remove duplicates with findimagedupes, handcheck for quality, and then upcale the remaining ~6000 to 512px with Waifu2x.

I began training OpenAI's Glow and Nvidia's ProGAN on it; the ProGAN is going very nicely, and I didn't need any data augmentation, but if I do, it'll expand the dataset even further.

So that might be a good strategy for MakeGirlsMoe: create something like a body locator, extract a few million samples from Danbooru2017, use Waifu2x to upscale to target size as necessary, ???, profit.

[DemonRem]

thank you Aixile

[Burninator448]

If you need a large dataset for this I'd suggest you take a look into the Nico-Illust and Danbooru2017 datasets.

They both seem to be quite capable for GAN training data, as well they both have large amounts of metadata which should trivialize the task of sorting them out.