Adverb @AdverbAdjectiveNoun@refactorcamp.org

Losing in the Latency

#generativeart #mastoart #gif

Honestly, the dream is to find this but for glasses...

Found a single scalar in the latent vector that works pretty well for masc/fem appearance #ai #machinelearning

Next step is to find my own face in the latent space and alter it

(Extremely divorced 45-year-old man voice) The National's new album is practically out if you don't mind a video with it.

Not that I dislike the film portion - I'm just leaving it open in a different tab to hear the soundtrack.

This is made by setting the latent vector to be constant and the first layer of noise to change.

Surprisingly, the subsequent noise layers do very little aside from micro-details. This may be because I didn't regularize the noise by switching it every now and again.

I've written what is essentially a StyleGAN-Light, I think. #ml #gan #ai

My model uses just over 8 million trainable weights compared to ~23 million in the generator alone for Nvidia's StyleGAN.

This is trained on a Kaggle kernel for less than 10 hours on the celebA dataset at 256x256 resolution.

Pretty smooth.

#gif

She wants to change

(celebA) #mastoart #generative

Moving houses

#gif #art #machinelearning

Here are the lower-level, more global changes I alluded to when you only alter the first noise-addition layer and the rest are left the same.

As you can see, it really looks like the same room, or a room of nearly identical style, but the walls have shifted whether by perspective or otherwise.

My GAN works very similarly to StyleGAN in that the fine details can be altered based on higher-level layer noise.
You can hardly see a difference here, but wait until you see what the lower-level noise does.

(As an aside, I still need to decide a name for my GAN, which uses some elements of the Style- and Big- GANs but with much fewer parameters. For a while I was thinking REA-GAN, since it fucking sucked [like the president], but now I'm too happy with it.)

Description of images:
two slightly colored-in frameworks of white and brown that seem to suggest several plausible perspectives on a room

And here is a dreary motel room, for your viewing pleasure.

I've finally gotten to the point with this GAN where I can play around with noise to see how it affects fine and gross detail at different levels. Exciting!

Here are two images of it that seem to reveal an underlying default, barebones to the generated LSUN rooms

Getting some success with 10% of LSUN...

#ml #ai

Looking for a room.

#machinelearning #gan

The most interesting part for me is the 4,1 and 5,2 images where the barebones black and white structure seem to be shown.

I feel like I could get a rock to generate good 128x128 images, but lately set it to 256 and all hell breaks loose :,)