Type | Other |
Stats | 711 |
Reviews | (352) |
Published | May 8, 2023 |
Base Model | |
Hash | AutoV2 865451DD46 |
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://civitai.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.