The one with the creepy eyes and the broken legs and the best drag show this dynasty has ever seen.
Making
As part of my deal with my friendly local supercomputing cluster, I got a couple free hours on their A100 server-grade GPUs. These are the sort of GPUs that cost more than a car, the kind that people use to train and/or run large langle mangle models on. I have a couple modest consumer graphics cards, but my best one is an AMD which isn't really in vogue for the hot new stuff. So I sat down and thought: how best can I waste this A100's time?
My trial run was sam-3d-objects, a set of models designed to generate a 3d model from a single 2d photo. You see, traditional methods require many 2d photos due to the nature of 3d objects to have a backside that a single photo cannot capture. But someone brave at facebook said "what if we just… guessed what the backside looks like?" and now we have a model that can take any single photo and guess the overall geometry, including any obscured features. I thought this was kinda neat, mostly useful as a replacement for quickly generating models for furniture/appliances in my home but potentially useful for 3d-ifying images for blender shitposts in the future too.
I've used plenty of python packages before, usually by throwing a requirements file at pip or uv and then importing the function that does what I want. This week I learned that ML python is a whole different beast. First you have to get the model weights, which are technically restricted by trade laws but I just had to hand over my email to facebook and get signed up to their oculus marketing emails which apparently proves I'm a red-blooded american. Next, you have to build the container image following the steps listed in their docs, which includes a couple hacks, patches, and overridden index URLs. This takes thirty minutes and fails because some dependencies don't work well with build isolation, some need to be built from scratch, some need even more patches, etc. Each subsequent rebuild takes anywhere from thirty to seventy minutes, and when you finally have a working container you get a dozen warnings about deprecated functions, outdated modules, and things that may or may not be broken depending on the phase of the moon.
THEN, you have to actually try to use the library. This repo is a proof of concept, an academic demonstration, not a library fit for the masses. You will find a notebook with a nice example, as a courtesy for those who have made it this far, which contains a few dozen lines of very simple, very reasonable code. The impression is that this is all so perfectly intuitive that it needs no explanation. And then you look at the actual imported python script and it's a mess. This isn't a library at all! They've hacked their own shit into a dozen pieces and plastered over the cracks to make the demo notebooks seem in any way reasonable. This might be normal for ML python but it's insane how bad this is compared to any reasonable modern software.
Anyway, I got it running. By this point I had forgotten why I started this project in the first place and instead plugged Toast into the Algorithm, which gave me this:
It's a little rough, particularly around the face. I think the whiskers threw it off a bit. But it is a full 3d model, based on a single image! They did the thing! They even got his fucked up back arch!
Facebook also released a model and whole other pipeline specifically for bodies, which I imagine would handle faces a little better. But I paid no heed and kept using this model, throwing Ginger and a local stray (Fancy Bastard, sworn enemy of Toast and Ginger) into the mix. I also threw in our Christmas tree to see how detailed it could get the resolution (not very).
So yeah, not great at faces, not great at resolution. Later I tried it on some houses and it worked much better. Big blocky things tend to work best, but those are also the easiest to model by hand so this thing probably isn't the biggest development since canned bread. Very funny though!
We also played two games of Heart: The City Beneath. One was a trial run, which was great fun fucking around. And then the actual first game of the campaign in that system, which went quite well and will save me much stress in the future. Good system overall - not perfect, but quite good.
Reading
I read a good chunk of Baru last night, which was nice. It was a little uncanny to read the scene of the head of state threatening the Imperial Accountant with bodily harm, then waking up this morning to the news.
I am extremely interested in Baru's character. She seems to be the person taking things more seriously than anyone else ever, which will either lead to her winning everything or absolute ruin. There's also like seven layers to her mask at this point and we've looped back around to the earlier ones.
Watching
I finished Edgerunners, which was well done! I liked that the show dealt more with cyberpsychosis than the game, but neither did very much all said. Cyberpsychosis is one of those blatant game design cudgels, an obvious downside to the cool-as-hell body modifications and upgrades the game offers you. And in the show it's the natural consequence to characters getting too much power, which becomes a kind of trauma very entwined with their capacity and propensity for violence. But it's still not particularly convincing as an outcome that follows from its causes, or as a thing that would continue to exist in the world where supersoldiers who can continue to follow orders are wildly valuable.
All said, I think Edgerunners is good but hard to recommend. The show did a good job of echoing the mood of Night City, the ruthlessness of the corporations, the feel of playing a character in that genre where you feel the need to make your mark in a spectacular way and destroy yourself in the process. In that sense I think the show did better than the game in many ways. But it is a show centered around violence, nudity, and self-destruction, and glorifying those things in ways that are harder to ignore than the game.
Playing
I played a little bit more Witcher Tres this week, mainly exploring the south of Veden. Still haven't done any MSQ since arriving, but I do need to hop on that soon so I can dispel illusions.
turns out, if you put whole milk yogurt in with the milk when steaming it, the texture ends up being PHENOMENAL. Cloudy, fluffy, and the foam holds for a while :D it's like more milk per milk?
