Aitrepreneur
patreon
STYLE LORA Training Presets Kohya
🕑 Added 2023-09-10 17:48:39 +0000 UTCHere you will find 5 optimized training presets for STYLE LORA inside Kohya SS. Whether you're training locally or using Runpod.
What's inside?
3 Local Training Presets: Designed for those who love to train on their own systems.
- Normal Style Training: For your everyday needs.
- Fast Style Training: For those in a hurry
- LowVRAM: Optimized for local systems with limited VRAM.
2 Runpod Presets: Perfect for those using Runpod's GPU renting platform. These presets automatically locate the location of the stable diffusion XL model for you.
- Normal Style Training: The classic training preset.
- Fast Style Training: Train your models as fast as possible
Why These Presets?
For LORA training, every little optimization can make a huge difference. These presets are the culmination of countless hours of experimentation and refinement. By using them, you're ensuring that your LORA STYLE training is both efficient and effective, regardless of where or what you're training.
And as always, supporting me on Patreon allows me to keep creating helpful resources like this for the AI art community. Thank you for your support - now go train some awesome LORAs!
πDownload Style Presetsπ
password: PRESETSTYLE
Comments
Ricardo Campos
it works! my fault! Thanks for all your work
Aitrepreneur
Are you sure you're doing things right? Look at this: https://imgur.com/a/OAdhgsO First upload the preset file inside the kohya folder, right click on the file, select "copy path" then inside kohya, paste the path, DO NOT FORGET to put a / before the path, then click the load button
Ricardo Campos
Daniel Alderson Smith
I have been comparing the different training presets you put out, using them as reference to try to get an undertanding of the network rank and alpha settings... You have for the SDXL network rank of 256 and alpha of 1, but for the sdxl low vram u have rank 24 and alpha 12.... but then for the style presets, the low vram has rank 24 alpha 1, and reguar SDXL-style has rank 256 and alpha 1... could u help me understand why you set them up like this. alpha of 1 for all except SDXL-low vram?
Aitrepreneur
you can use them as is, it's the same parameters except the model change (untick sdxl model checkbox) and change the reolution training from 1024,1024 to 512,512
Xavier LarrodΓ©
Hello, do you have a version for SD1.5 ? Thanks
Aitrepreneur
what json file? A lora is a safetensors format
BadgerWi Johnson
I have a problem the .json file that I created in Kohay is not being seen in stable diffusion when I go to look at and load my loras. Loras I got from you or others I grabbed from CIVITAI show up just not he one I created in Kohay.
Aitrepreneur
emotions yes but nothing that is part of her face, unless she wears different colors lenses on each photos, if she always have blue eyes, leave it be
DAVID OSTLER
I'm mid way through your LoRA tutorial, the auto caption described " a woman smiling" Should I include things like smile, blue eyes, or emotions in the caption?
Aitrepreneur
that's a really good resolution, that's also what I usually use, but no need for anything bigger though
Remgrandt
What is the image size I should aim for with my training images? I have some that over 4,000x4,000, is that too big?
Aitrepreneur
you can increase the batch size for faster training
Hi! is there any settings (SDXL character) that I can use to train faster without reducing quality? i have 24gb of vram. Thanks!!
Aitrepreneur
Send me screens on discord
Can i ask you a question? My LORAs generate lot of 'confetti" portraits. Like if Picasso decided to to paint my photos. What may be causing it?
