• mine craft! mc.floofey.dog:25565

Archive

  • Nobie @ Nobie:
    What does it take to train your own llm?
    Quote Link
  • L (Guest) Lexus:
    site is cool, maybe one day i'll come back. stay safe, happy new year
    Quote Link
  • Adel @ Adel:
    llama didn't really
    spark my interest
    Quote Link
  • Xeraser @ Xeraser:
    not really, i don't dabble in llms but i wanted to get llama 33b running for shits and giggles
    Quote Link
  • Adel @ Adel:
    I've been looking for llms in general, know any cool ones off the top of your head?
    Quote Link
  • Xeraser @ Xeraser:
    i'd wager most useful llms would take 8-12gbs
    Quote Link
  • Adel @ Adel:
    and iirc it didn't ask for too much in terms of hardware, mainly VRAM
    Quote Link
  • Adel @ Adel:
    I recall it ran locally
    Quote Link
  • Adel @ Adel:
    there was one which was primarily intended for creative writing or something along those lines
    Quote Link
  • Adel @ Adel:
    speaking of chat models, I remember browsing a few offline ones a year back but I can't remember the name for the life of me
    Quote Link
  • Adel @ Adel:
    I recall when chatgpt was opened to public, super breakable
    Quote Link
  • Xeraser @ Xeraser:
    more or less
    Quote Link
  • Xeraser @ Xeraser:
    december '22 to march '23
    Quote Link
  • Adel @ Adel:
    since that was the most directly useful thing to me
    Quote Link
  • Xeraser @ Xeraser:
    the big boom was last year at around this time
    Quote Link
  • Adel @ Adel:
    mostly with upscaling
    Quote Link
  • Adel @ Adel:
    but yeah, as per your previous comment, I mainly kept up with AI stuff around 2017 through 2021 ish
    Quote Link
  • Xeraser @ Xeraser:
    i just moved on to NAI v3
    Quote Link
  • Xeraser @ Xeraser:
    i should update the SD thread/guide whenever i have a local install at my disposal
    Quote Link
  • Xeraser @ Xeraser:
    the included models in waifu2x-caffe have been enough for asset upscaling with some touchups
    Quote Link
  • Adel @ Adel:
    upscaling's super fleshed out
    Quote Link
  • Adel @ Adel:
    and every artstyle of anime
    Quote Link
  • Adel @ Adel:
    along with all the animes
    Quote Link
  • Adel @ Adel:
    yeah, generally upscaling video game assets
    Quote Link
  • Xeraser @ Xeraser:
    there's already remacri, animesharp, ultrasharp, ultramix, lollypop, etc
    Quote Link
  • Xeraser @ Xeraser:
    upscaling models? like esrgan models? what for?
    Quote Link
  • Adel @ Adel:
    though, most of them are training upscaling models
    Quote Link
  • Xeraser @ Xeraser:
    i'll be very honest with you fam it sounds like you tried it for maybe a couple hours like a year ago and never looked into it further
    Quote Link
  • Adel @ Adel:
    I was under the impression that it took longer, judging from the complaining I hear from all these nerds
    Quote Link
  • Xeraser @ Xeraser:
    less if you just autotag
    Quote Link
  • Xeraser @ Xeraser:
    Adel said:
    let me elaborate, you can have all the control, however the time it takes to train a model on a custom dataset which has the style you want to replicate is better spent learning to draw
    it takes as little as two hours between curating the dataset, tagging and training
    Quote Link
  • Adel @ Adel:
    then again, none of it is truly original
    Quote Link
  • Adel @ Adel:
    LoRA exists and you can mix stuff
    Quote Link
  • Adel @ Adel:
    actually, that's not entirely true
    Quote Link
  • Adel @ Adel:
    you don't have any control over other models
    Quote Link
  • Xeraser @ Xeraser:
    uh
    Quote Link
  • Adel @ Adel:
    Xeraser said:
    you have A TON of control
    let me elaborate, you can have all the control, however the time it takes to train a model on a custom dataset which has the style you want to replicate is better spent learning to draw
    Quote Link
  • Xeraser @ Xeraser:
    Adel said:
    guys will sit for literal hours at a keyboard and type up prompts just to get a somewhat decent result
    massive skill issue tbh
    Quote Link
  • Nobie @ Nobie:
    I guess I can see where you're coming from but at the same time we never had anything like this EVER and it's pushing the bounds so I'm just happy that is what it is
    Quote Link
  • Xeraser @ Xeraser:
    Adel said:
    also it's not better since you don't have much control at all
    you have A TON of control
    Quote Link
  • Xeraser @ Xeraser:
    that's because you have to put SOME
    effort into it
    Quote Link
  • Adel @ Adel:
    give it another 5 years and we should have better stuff
    Quote Link
  • Adel @ Adel:
    guys will sit for literal hours at a keyboard and type up prompts just to get a somewhat decent result
    Quote Link
  • Adel @ Adel:
    the good results you see make up literally 0.01% of all the generated results
    Quote Link
  • Adel @ Adel:
    as someone who dabbles in the stuff, it's really limited
    Quote Link
  • Adel @ Adel:
    also it's not better since you don't have much control at all
    Quote Link
  • Adel @ Adel:
    it's useful as a snapshot for getting people acquainted with the stuff
    Quote Link
  • Nobie @ Nobie:
    Image AI is better than actual artists at this point I'm convinced(At least for anime drawings)
    Quote Link
  • Adel @ Adel:
    the basic stuff which explains the modus operandi still stands, but some of the tools I used are definitely out of date
    Quote Link
  • Nobie @ Nobie:
    AI evolves quickly
    Quote Link
Back
Top