FLUX UPDATE !!!
I decided to update my Lora for Flux, at 4k steps and 3k steps.
The preview images were the outputs I got when training at 512 x 512 on 29 images.
I do believe the 2000 and 2500 step models were also pretty good, should I upload them as well ?
I hope you enjoy :)
Introduction
This is the first LORA I've ever made, so it's far from perfect. I trained it on 85 images of Bacun's art using Anything v4.5 at 512x512.
In hindsight, I realise some of the images I included in the training set messed with the results, as they have multiple characters or wrong anatomy. Five of them didn't even get picked up by the Kohya SS gui. I guess that proves quantity doesn't always equate to quality. Also, sometimes it can get pretty close to the referenced illutrations.
Usage:
For now, the LORA needs heavy handholding.
I'd say put the strength at 0.6 - 0.8 (over those values, impurities and other characters will often pop up). If you enable highres, put the Upscale at 1.5 and Denoise at 0.5
This combines quite well with other character and style LORAs, you can also try using ControlNET.
What's next:
I'll try to improve this LORA some more by polishing the training set, then the next LORAs I'll make will be one focused on a specific Niji Journey cartoon style, and one about the character Arezu from Pokemon Legends Arceus.
As for why I made this, it was because I wanted to learn this artstyle, so I thought making a LORA based on his art could be a huge help with learning the anatomy. I was gonna keep it private, but changed my mind in the spirit of open source. Also, it would only take a few weeks or months for someone else to do it instead sooo yeah.
Final points:
Please, if you make something with this model, credit the original artist (Bacun), as he's a very talented individual. Feel free to polish it, use it in mixes or whatever.
Feel free to use it locally or on Google Colab.
Please, do not use it on your generation service without my agreement first (although I obviously can't enforce it).
I hope I don't end up regretting posting this, but hey, it was interesting to discover the inner workings of LORA-making. This shows that literally anyone can do this after a couple of hours though.
Edit: Removed early access, thought it wasn't necessary in the end.
Description
First version, it's far from perfect so I need to clean the training set
FAQ
Comments (17)
Ah you beat me to it. I'll need to test it out later. Congrats on your first LoRA!
Thanks, it's pretty scuffed for now but I think I can optimise the dataset.
It would be great to have an insight from better LORA makers so I could figure out how to push it to the next level.
But yeah, if you're experienced and already prepared your own dataset, you can easily do better than this considering some of the references have used had weird anatomies, multiple characters and that I kinda speedran tagging all 85 images.
Also, your Metata style LORA is really cool !
@SomeAIGuy I don't have the dataset yet or anything. It's just in my backlog. I have a lot planned but can't decide on an order ever haha. Are you using Grabber to get the images by the way? It can save a lot of time by getting the tags when you get the images at the same time. Then you just need to make whatever edits you's like. Also thanks! I'm glad you like it!!
@thesauceer As said in the description, I made this LORA because I thought it could help me understand anatomy/shapes more easily. Because I was studying his art, I already had most of his illustrations downloaded (one by one). I just added some more recent ones.
Individually editing the images, then using deepbooru to tag them one by one, and then editing the text took me a very long time, I went to bed at 4am yesterday (and that's before early troubleshooting which took me the entire day).
If I had more vram, I could have trained at 768x, but hey, it's good enough for now :)
@SomeAIGuy Ah gotcha. Good luck in your art endeavors!
@thesauceer Actually, where can I get the Grabber thing ?
@SomeAIGuy Make sure to go through the settings and change it so it grabs the tags as txt files too! https://github.com/Bionus/imgbrd-grabber
@thesauceer How can I do that exactly ? (like what setting do I have to change)
@SomeAIGuy To be honest I do not remember. I can take a look at it on thursday when I get home and let you know though.
@SomeAIGuy I have mine set up like this, but you'll need to change the folder structure and potentially not use %all% for the tags you want.
Options are under ctrl+p by the way
@thesauceer Thanks, I figured it out. I think I'm missing something though since the only source that makes images appear is Deviantart (the rest is "offline" or something). Are there any guides online ?
@SomeAIGuy Here's the official docs.
All the provided examples use the LORA as it is, I'd like to make a mix with @Nerfgun3's Yaro Artstyle LORA, but I'm pretty sure I need his permission.
Then I'd also have to figure out how to do it. But yeah, try generating using this LORA combined with his, it works amazingly well !
To give you an example, I added 3 example mixing both.
You don't need nerfgun's permission to merge with his model. According to the license you only need to give him credit.
good job mage likes it
The style of 85 pictures is very clear. Have you made any changes on the tag?