CivArchive
    [Guide] How to train MORE LORA models in LESS time and in a simpler way? + Free Notion template [ESP/ENG] - V1.0
    NSFW
    Preview 996245

    UPDATE 6/20/25: THIS MODEL MIGHT DISAPPEAR IN THE NEXT FEW WEEKS! More information at: https://civarchive.com/models/1660755/list-my-shadowban-models-13-and-counting


    PLEASE READ: the guide is not the file that I am passing, it is in the link

    La guía no está en el archivo, sino en el link de abajo.

    I am pleased to announce the milestone of the creation of the first major LORA modeling protocol, a protocol that will help us reach that latent unexplored space of limitless creativity faster.

    Notice: Check the articles section in the future, as I will end up moving to that area,

    • Protocol problems; Depends on Google colab in part

    Description

    FAQ

    Comments (16)

    fitCorderJun 1, 2023· 2 reactions
    CivitAI

    Can you make it so we can read it on civitai?

    TecnoIA
    Author
    Jun 1, 2023

    There are reasons not to do it like this, it would mean making a post with a giant size in the first place by including both English and Spanish, it is easier to separate by language into different posts

    malcolmreyJun 1, 2023· 22 reactions
    CivitAI

    the title is a little clickbaity, because there is no technical "revolution" that makes you train LORAs faster

    it just talks about being organized (which is good)

    I will however comment on certain suggestions in that guide:

    Perfect preview photos: I recommend adding between 3 to 6 very different-looking photos

    If you are lazy, yes, 3-6 is fine :-)

    I go for the max which is 20 for the following reasons:

    you can showcase your model in a variety of prompts AS WELL as a variety of models.

    This shows the users that your model is consistent over different base models and works well in various scenarios.

    Is that LORA useful or necessary for the community in general or just a whim for yourself? PD: There's nothing wrong with doing strange or unpopular LORAs, but it's helpful to prioritize the ones you know will have high demand.

    this is a tricky area, sometimes when I upload something I think, this will be popular but it isn't and on the other hand, I upload someone and I ask myself in the voice of Michael Bluth (Arrested Development) - "her"?"

    This is the approach I recommend. It is recommended to do it in separate sessions for each phase, for example, one day you plan to create several LORAs, another day you create the dataset for each one of them, another day you train them. It is recommended to do 6 or 7 LORAs per phase with this method. (Later on, I will specify the phases and the steps to optimize each phase)

    this is actually quite good advice :-)

    at this moment I have like 200 downloaded possible subjects, and around 20 models where I've already picked ~40 best pictures, like 10 models where those pictures are already recut, and 5-10 models where I've filtered out the bad results

    there are several stages of data preparation and it is indeed more efficient to focus one day on downloading images of new subjects, on another day on preselecting the good images, and on another day - on cutting them to proper sizes, etc.

    if I was to do one model at a time, it would be actually less efficient

    Making up to 12 quality LORAs per week without wearing ourselves out (at least individually).

    I can paste this meme here:

    https://imgflip.com/s/meme/Monkey-Puppet.jpg

    with the text "well, this is me who sometimes does 12 a day :P"

    making samples and uploading them,that's a different story :-)

    overall, good read!

    TecnoIA
    Author
    Jun 2, 2023· 1 reaction

    I like sensationalism, I was initially going to put protocol for LORAs on it but it sounded boring to me, thanks for the comment, I don't recommend doing so many LORAs a day huh, from experience it tires a lot (Today I did 12 like you XD)

    malcolmreyJun 3, 2023· 1 reaction

    @TecnoIA 12 per day, everyday - that would be indeed overkill

    cheers :)

    royal_dreamerOct 1, 2023· 1 reaction

    Thank you for your summary.

    ozm333Jun 1, 2023· 4 reactions
    CivitAI

    All of these "how to create LORAs" never dive into the training parameters, anyone have a link to a guide that dives into that? I wish my pc was strong enough to test every parameter out but that's not the case

    TecnoIA
    Author
    Jun 2, 2023

    I leave a link to a more technical and simple guide at the beginning if I remember correctly

    ozm333Jun 2, 2023

    @TecnoIA  I've read that one before and its a good basic but its not very in depth plus its outdated, now there are more options

    TecnoIA
    Author
    Jun 2, 2023· 3 reactions

    @ozm333 So I've been outdated equally xD, I join your request

    skavenger0Jun 13, 2023

    I found this guide an excellent start https://www.youtube.com/watch?v=70H03cv57-o

    IntisuyayJun 8, 2023
    CivitAI

    Gracias!!!! le echare un vistazo :D

    segment800Jun 28, 2023
    CivitAI
    TecnoIA
    Author
    Jun 28, 2023

    Thanks for the notice, what language do you want?

    mogJul 26, 2023
    CivitAI

    sredni poradnik a propo zrob model lub lora bakugan

    royal_dreamerOct 1, 2023
    CivitAI

    Thank you very much for sharing. It's very detailed and helpful.