Diego @NeurIPSさんのプロフィール画像

Diego @NeurIPSさんのイラストまとめ


Guatemalan 🇬🇹 physicist, now Ph.D. student @CVC_UAB, researching autonomous driving.
blog.diegoporres.com

フォロー数:335 フォロワー数:952

Previously I used --downsample-factor=12, so setting it now to 8 greatly stabilizes the interpolation, but also increases the rendering time ~4x on an A40 :/ https://t.co/JgwzVDbDnN

0 5

"The Fermi paradox", left, right

There's something uncanny with the way both of these models interpret this simple prompt, but I enjoy the way a human is added by because, well, it's a paradox that only we care about in the end

0 3

"A dog with a fly's body", left, right

Note that I had to first generate a fly, then edit the face and change the prompt in order for to give the shown result. At least does something with the final prompt, but it looks unrealistic/synthetic

0 4

This is it, all my knowledge of matplotlib has culminated in this plot. I am betting seaborn could've done this easier, but something something sunk cost fallacy

1 10

It even works with the visualizer .py (forgive the cursed latent)

0 6

Somehow managed to import 's extended StyleGAN2 to my SGAN3 repo (https://t.co/6E2m2DUnmL). The idea is to use it for fine-tuning, so I'll compare it to the latest training iterations I've done (SGAN-XL, SGAN3, SGAN2).

3 12

Absolutely love this SIGGRAPH '20 paper: explanation of the illumination in Salvator Mundi especially the three white spots in the orb. The optical accuracy of the painting has cast doubt on the attribution to da Vinci, but this paper shows it's correct. https://t.co/UGCZmtbaih

0 5

Thanks! This one reminds me of the weird dogs in DD

Someone did do DD with CLIP and they showed me a weird Buddha that came out, but that tweet is lost. I'll try to look for it!

0 3

Test 2 of interpolating in (448x448):

Sometimes, it just misbehaves/the latent space is *too* choppy. Other times, it's super smooth. Perhaps it's worth to look at the attention maps/where the network is concentrating, and malleating this?

0 6

Some seeds produce weird images though, most likely coming from the Perlin noise. This, however, is a great way to tile an image, so will tinker with it...

0 4