CivArchive
    Aurora C - Aurora C
    NSFW
    Preview undefined
    Preview undefined
    Preview undefined
    Preview undefined
    Preview undefined
    Preview undefined

    A Russian model from back in the day. "Aurora C" on Stunning18, aka Tita from Nubiles, Petty for some hardcore stuff.
    This is very much a WIP. I can't figure out which version is best and may redo it, but here is what I have for now.

    Feedback is very much welcome.

    Description

    FAQ

    Comments (7)

    regajet336Feb 3, 2024· 3 reactions
    CivitAI

    Hey, good first try. Not overbaked and likeness is fine. Just some friendly advice to improve: get rid of the watermarks and use some "neutral" poses for training (like standing, sitting ... basicly avoid anything where the face is sideways or shes leaning, lying...) . You will notice after that your LORA is way more flexible and quality goes to the roof. Cheers mate :-)

    SoftNSquishy
    Author
    Feb 3, 2024· 1 reaction

    Hey, thanks a lot for the constructive feedback, really appreciate it! Good to know, I was wondering about postural issues too, will take on board when I get around to v2. Practice makes perfect I guess!

    regajet336Feb 3, 2024· 1 reaction

    @SoftNSquishy No Problem .You got a good taste and i hope i will see more of your Loras. I'm still learning to, but those were the mistakes that cost most of my hair and weeks in time to find out. Nobody tells you this in videos or tutorials. Only "bla bla watch shadows, details, high quality" and other stuff that in my opinion doesnt matter to much. If you like the style and quality of the source you get the source (garbage in = garbage out). Its really that easy. As for Poses; for example you will notice, that if your girl is leaning forward in most pictures you will get comicly big heads sometimes. Why? Because of the camera angle the body is smaller, so the ai asumes its okay like that. Same goes for lying / leaning / bending over images. Avoid them like the plague. The ai is stupid and cant handle that in training. In short. Neutral poses are better and try to use way more face (2/3) then body (1/3) Images. Another small Tip for Watermarks: just paint it over with black paint, it will be ignored in training ;-) IMHO best Tutorial: https://civitai.com/articles/1824/how-to-make-a-lora Hope i could help and happy creating. :-)

    SoftNSquishy
    Author
    Feb 3, 2024· 1 reaction

    @regajet336 Solid. I think the ratio of poses vs head/torso shots was definitely a factor. Regarding the watermarks, I believe I was simply sloppy in this case (as in one of my earlier LoRAs where I completely failed to remove all the onlyfans watermarks lol).
    Thanks for the link, there's a lot of bullshit and conflicting advice out there which, frankly, put me off even trying for quite some time.

    regajet336Feb 4, 2024· 1 reaction

    @SoftNSquishy Hehe yeah, i totaly get you. To get to that level of Quality you deliver, it takes painful weeks of trying, tinkering and frustration. Theres to much junk tutorials out there which will lead people in the wrong way or are overly complicated. Tryed dozens of Guides for months and still got garbage most of the time. Annoying as F***. One eye opener was the advice in the linked tutorial (making folders with same trigger word but different repeats depending on quality + get rid of the class trigger) and another one (sadly dont remember), which was the first to tell that you dont need to tag your characters likeness (like hair color). Just tag the things you(!) want to change. Speeding up the process of preparing greatly and gives waaay better results. Dont say its the holy "how to train stuff" advice, but at least thats whats working right now for my part. Still not sure about some things like Reg images... Did you use some?

    SoftNSquishy
    Author
    Feb 4, 2024

    @regajet336 Yeah all of that is good advice, basically what changed things for me at least! Nah, I haven't messed with regularization yet, it's on the TODO list somewhere 😂

    regajet336Feb 5, 2024

    @SoftNSquishy Hmm to bad. Could not find a clear answer if they are good or not. On self testing it wasnt clear for me either. Well, guess i have to tinker a little bit more. Until then cant wait for your next attempts :-)

    LORA
    SD 1.5

    Details

    Downloads
    216
    Platform
    CivitAI
    Platform Status
    Deleted
    Created
    2/2/2024
    Updated
    5/12/2026
    Deleted
    5/23/2025
    Trigger Words:
    4ur0r4

    Files

    325890_training_data.zip

    Mirrors

    CivitAI (1 mirrors)

    Aurora_C-000006.safetensors

    Mirrors

    CivitAI (1 mirrors)
    TensorFiles (1 mirrors)