Lazy-ass low-effort LoRA (LALELoRA?)
DISCLAIMER: I hate this way of making LoRAs. It is effective, it is fast, it is a mess. I can spot a few LoRAs on Civitai which were done this way.
This is how I made kCuteTanks, kCuteCreatures, kHorrorCrabs, kVoidEnergy, and kHastur.
It takes 5 minutes going from idea to clicking the "train" button.
REQUIREMENTS: Midjourney, Python, 5 minutes of time
Minute ONE (data selection)
Open Midjourney, select explore, write your idea (like "cute creature"), hit Enter.
Scroll down, click "expand results".
Click "enable selection" in the top right corner, then click "select all" in the bottom.
Optional: repeat with variations (like "cute animal")
Minute TWO (data collection)
Click "Downloader" and download all. Unzip all files to a single folder.
Wait a minute -- we didn't even check the pictures. What if there is a picture which does not fit the idea?
First, we can't wait a minute. We have to make it under five.
Second, do not care, this is the lazy-ass low-effort approach.
Minute THREE (captioning)
The names of the downloaded files have the following form:
[user]_[partof_prompt]_[id].png
Guess what: we just take the middle part and make it a caption.
We can also add a trigger word, and remove some words, like "cute", "creature".
GPT can easily write a procedure, which will do it for you. Since Civitai needs a file uploaded, I'm putting a python procedure I used.
Wait a minute -- The procedure does not ensure removal of the complete username. Also, there is only a part of the prompt. Such captions will be a mess!
First, we can't wait a minute. We have to make it under five.
Second, do not care, this is the lazy-ass low-effort approach.
Minute FOUR (training parameters)
Just take the config for kohya-ss from the other tutorial (https://civarchive.com/models/52697/tutorial-konyconi-style-lora), set the number of epochs to 1-4 (depending on the amount of pictures), and click the "train" button.
Wait a minute -- Shouldn't we fiddle with the parameters, find what is best for the specific LoRA we are making?
First, we can't wait a minute. We have to make it under five.
Second, do not care, this is the lazy-ass low-effort approach.
Minute FIVE
Oh, there is no minute five. We actually could wait a minute.
Description
FAQ
Comments (32)
Going back to experiments with TIs.
I've had a go with your Boho data. Does Boho'ish stuff :-)
@chromesun Great!
I like lazy, asses, low effort, and loras. This must be for me!
Well... not surprise, haha since settings are always secondary on training and images are the most important input, even you let username there it will work if the images are good quality enough. Just the trigger will act as a bleed between the +500 images you want to drop and 128/128 will act as a maximizer.
If you ask me...better curate the data. I'm releasing also an approach to Lora lazyness (similar to my turbant) but pointed to inferencing + doing evals. Just writting male/female will spit painterly outcomes (using my own digital artworks), towards the Lora training, etc. still tweaking DIM/Alpha. Wrote a post in Reddit about this in case someone wants to check.
Btw, thanks for the python script
Thanks for the feedback. I completely agree, that it's better to curate the data (both pics and captions). This was to represent the lazy-ass approach.
I've honestly read everything on these posts and I'm just as clueless as when it started. ima need an updated youtube vid, lol.
@SomebodyOnceToldMeh Um, you mean on the Kony process? of what I said? lol, now im the one confused.
HEY this is actually TO. BE HONEST - Worth the 10 bucks to do MJ for - When you KNOW what you're looking for. This isn't EXTREMELY lazy, I actually salted some MJ stuff into a couple of lora recently. I can't train locally, but hey GOOD TUTORIAL!
You can't train locally, so what do you use to train? A collab?
@Servo_Scribe Yea I use colab specifically Holo's one laetly.
@duskfallcrew cool, where do you find that?
@duskfallcrew Cheers! So I need to upload my folder of pics/captions to Drive first eh? (Also, would you happen to know how to implement the Python script? I think I'm missing something
maybe someday ai could do this by itself
man, I think I could program a single python file to do all this, or better, ask chatGPT to program it lmao
@VCominos I can't even figure out how to do use the python in this tutorial lol. I keep getting "IndentationError: unexpected indent" lol
Where exactly do you click explore on Midjourney? Couldn't find it on both Discord and their website.
It's on the website: https://www.midjourney.com/app/ ... in the menu on the left, it is the second item from the top.
I think Only subscribe members can see the Explore Thing!
I can't find this because i'm not subscribe to the plan
At this point you're making it seem so easy it really is giving me no reason not to try it. Guess it's time to re-up my MJ account!
Short question.
For 12 - images of the same style. Not random styles like your datasets.
How would you set the settings up ?
Trainbatch: 1
Epoch 6 ? or 15 ?
repeats: 10 or more ?
cousine or constant ?
if i forgot something, maybe you can add it :D
Because i have the problem, that those small databases with one style arent working very well...
Sadly...
Any tips ?
And bro, amazing work!
I bought a 3060 because of you :D
My RX6800 laying in the corner right now xD
Honestly, I've never tried that. I think you should get 1500-2000 steps. My first try would be 150 repeats, 1 epoch, constant. @LDWorksDavid , what would you recommend?
@konyconi Thx for replying bro!
I will try it out :b
Btw. i achieved some cool diamond lora with your settings ^^
@konyconi Will post it under your other thread
@konyconi sweet spot general is 1700-2300 steps but to be fair each training is different. For example some works with 1000 and others till 4000/5000 don't start showing aesthetics... for constant its recommended not going full 128/128 since that scales a lot and it's not act as factoring as cosine so it's linear (more or less), think of it as less variable. Maybe 64/32 its better choice or even 16/16 but 128 it's a lot for constant (of course dependant of learning rate too). Maybe 1-2 epoch, Batch Size 2-3. and repeats probably 100 or even 70 (150 maybe overfits? Or maybe no, I don't know).
Small datasets seems to be a bit more complex to train since AI tend to overfit or undertrain easier so probably best way is to get 2000 steps roughly and see how it went. From them lower learning rate or even Rank(Dim.
I wish I could give you a clear formula but is not possible, best take also is input images >>>> training settings (to a degree of course).
Im super lazyass 3,
anyone any question?
Damn, and here I though meticulously tagging and fiddling with settings, waiting a few hours for it to complete, testing the epochs only to find that none are as good as you wanted and throwing it out was the best way.
I need to try this new method!
Cheers! I'm a total noob and having trouble running the python script. Is there any context I might be missing as someone with no python knowledge? (if there are parts I shouldn't copy-paste into the python console, etc)
Open the script in a text editor.
Replace folder_path with your image directory path.
Replace triggerword with your chosen word.
Save the script as a mj_go.py file.
Open a command prompt, navigate to the script's directory.
Run the script with the command: python mj_go.py.
@konyconi Thank you! I just figured it out earlier today actually haha (just had ChatGPT hold my hand through it), turns out I'd been thinking the python console thing was a command prompt or whatever. I hope someone else can learn from my incompetence :P hehe