Pinned ping

that's the largest model too

the quality is definitely a lot worse than the training images but I think it still goes to show how there's a lot of redundancy in a light field

Show thread

what's kinda interesting is all the training data for the neural network is ~15 MB, and the code to run the neural network on shadertoy compresses to 50 KB

Show thread

experimenting with training the normal map and the colour simultaneously

training with colour now
early iterations with high loss look kinda beautiful

Show thread

this morning I am training a neural network to reproduce a light field

it looks very strange when asked to render outside the training viewpoint

every time I interact with someone on here who also uses it/its my heart glows

Show thread

I think it's good to remember that language is ultimately a form of data transmission, so you'll never get 0 perplexity on a sentence, there's a minimum number of bits which represents the information that is being expressed

Show thread

if you're curious what I'm talking about this paper does a good job explaining what text degeneration is arxiv.org/pdf/1904.09751.pdf

Show thread

idk who needs to hear this, but neural text degeneration happens because maximizing likelihood of a generated sentence is equivalent to minimizing entropy. so you're basically asking for generated text that contains *no* information

I made another music mix, this time featuring hand-drawn animations for each song!

shout out to blender's "import images as planes" which was instrumental in the animating process

youtu.be/WOAe-6ksObM

Show more
Cybrespace

cybrespace: the social hub of the information superhighway jack in to the mastodon fediverse today and surf the dataflow through our cybrepunk, slightly glitchy web portal support us on patreon or liberapay!