So, to start out I’ll be the first to admit that I used to be very cynical and dismissive of GenAI tools, but since i embarked on my learning adventure last year trying to make a digital avatar from a still image and voice sample my opinion has come around.
Sure, most people don’t really utilize it for anything of value, but the technology has gotten so far in the last year. I’ve taken to vibe coding my own AI tools using open source models. the landscape moves so fast and the coding tools are working so well nowadays that they can achieve my goals much faster than i can learn the syntax and how to navigate all of the dependencies required.
I’ve been obsessively researching and working on my wasp biology project and i’ve been building a machine learning pipeline for characterizing the near infrared spectra of the various pheromones on the surface of the individual wasps.
in addition to that, i’m building my own offline coding environment that i can use to analyze and modify any code i need in order to understand it better. So far, i can get it to output good lua and C++, but i hope to make it programming language agnostic in the future. it can currently utilize ghidra to decompile and analyze all manner of things and i’ve even used it to reverse engineer a couple of save files for my son’s video games.
I’m also working on a voice cloning and translation pipeline for my daughter’s birthday. She is a big fan of anime and just wants to watch shows with an english dub. so far i’ve had pretty good success.
while the LLMs might be good at coding, they tend to lack understanding and they definitely require tight instructions and oversight in order to get what you want. I’ve even had ongoing arguments with them at times due to just a poor understanding of the actual software engineering and architecture. However, once you learn how to speak to them it is an incredible resource.
What has everyone else’s experience been like? I hear a lot of doom and gloom, and a lot of frivolous and trivial uses, but has anyone out there actually put these thing to work in a meaningful way?
i see it as a new shiny toy. However i prefer to play with quallity toys. I was one of the early adopters of midjourney (image creation) and was able to get it ptoducing what i want as challanges back when it was still refusing to do what you wanted (version 3)
One of the developers there even told me, sometimes they fear my messages, because i find bugs in record time. Setting their neurons on the equivalent of drugs was my sugestion and created the parameter –weird
However i will never call any of those images AI, because computers can’t create art, even when photoshop users claim different. Set my mind to that back when i was a teenager and never saw a reason to change that afterwards.
Jailbroke the LLMs often just to see if i could do that. But i mainly use those as search tools and great statistic tools.
Also generated a few 3D assets because they got nice meshes, but that is not as much fun as breaking the other two.
In general, i just love messing around with AI of all kinds and figure out how to go beyond the intended usecases
@genesis have you had any luck generating consistent characters across multiple images? I’ve tried a bit with image stuff, but rarely been satisfied with the results.
@foggygoofball yeah that is one of the easy tasks. One if my favourite games was going into a public discord chanel of midjourney and plasting my face all over the place. Son Goku? Now me. Ork, defendly me! Sailormoon, yeah i saw worse, now a bearded dude looking like me!
The easyest was is using multiple reference images. The only thing those images should have in common are the character. It works better if the images have different styles, poses, backgrounds and expresions. That way the AI focuses on the stuff all images have in common. That should be only your character. Maybe the style, depending on your usecase.
Still use this old technique, however there are now specific character reference options in midjourney.
And it always depends on your character, what makes that character unique? What is the style? What are the identifyers?
Well frankly I’m not much of an artist myself and I don’t have the money to pay someone to draw a bunch of instances of the same character for me, I’ve had some success with Google Gemini and it’s Nano banana. It can generate almost the same image of the same person each time but the embellishments in their outfit keep changing from generation to generation. With my present experience and wrangling various AI tools into one coherent pipeline I think I might be able to isolate the elements and make sure that they all exist correctly. But that is a problem for the future, currently I am still working hard on my daughter’s birthday present.
Maybe next month.
@foggygoofball send me a message on doscord mabus_tiefsee i generate you some stuff for your daughter with Midjourney.
