I'm messing around with style transfer. Give me some images and I'll blend them

Messing around with pytorch

Attached: cars.png (851x643, 1.18M)

huh this thot photo has 4 channels and it made my model mad.

Attached: soviet.png (851x846, 1.51M)

Attached: cringe.jpg (262x234, 13K)

Attached: animu.png (851x491, 895K)

Attached: chuck.png (347x309, 174K)

Ty! Great work

Attached: CUM.png (1280x853, 58K)

What program or app is it?

I coded it up with pytorch.

Attached: woj.png (725x856, 1.12M)

Yeah thanx i found it

Oh ye of little faith. It's actually pretty easy to do.

If you're using a version you found online they're probably using the gram matrix approach from arxiv.org/pdf/1508.06576.pdf. I've been dissatisfied with that approach because images that look really different can appear identical by that metric. I'm experimenting with a variant based a Wasserstein metric instead. I think it gives superior results.

Attached: flag.png (851x576, 617K)

Looks interesting kinda, but yeah a bit repetitive

Humans tend to show less expected results

The details transferred depend on the textures of the style you imprint. The challenge is more transferring definitive style while still maintaining the structure of the image. Pic related is what happens with the old approach. It's hard to get a lot of style without washing out the image.

Attached: f564e7fe53b1615041e30c19a0e40813.thumb.jpg (3158x1000, 486K)

Needs more TPB. Explode the gif of Lahey falling down stairs, apply the transfer from a bottle of Hennesy and then stich em back together.

I believe in you Rick

Attached: Trump approves.jpg (746x953, 152K)

Attached: guy.jpg (729x969, 236K)

YOOOOOO

Attached: f95d7cea80c9be4bcb86eb922a2f688083677a291458f57175749aad4441e0ad.jpg (375x563, 34K)

Attached: 14358993_225849084484938_6300317000126214556_n.jpg (589x591, 46K)

make hillary clinton look like pepe