Credentials:
Introduction
I've officially reached 2 years on the platform and over that time I've posted thousands of models and burnt millions of buzz training attempts that will never see the light of day. During that process I found out what does and does not work and developed a workflow to assure me excellent results as long as that protocol is followed.
My goal when making these models is to have the highest fidelity/accuracy to my targeted style, model, or character. If a model looks amazing but does not match the source material I mark it as a failure and it should be retrained.
In the last month I finally left the experimental training methods and constant tuning I've done over the years behind and hammered down the settings and prep that produce the most recent models you see on my page. I can go from twitter profile to submitting a labelled dataset for training in ~5 minutes and you can too if you follow this guide.
This guide will pertain to gathering a dataset, tagging/labeling, and the training settings. I specialize in Illustrious/SDXL models trained on the Kohya engine but a lot of the principles are universal (ie. creating and labeling datasets) so I hope everyone can get something out of this.
This article will pertain to style models in particular since that's the workflow I have the most mastery over, I might update to include concepts and characters if there is interest, let me know.
Read, don't skim. I kept things very simple.
Step 1 - Crafting your dataset (40% of a model's grade)
The first (and to me the most important) step to creating a model is gathering the materials you plan to train it on. The easiest way is to find a database/catalogue containing a group of works and pull from there.
Websites I use:
Twitter -āāā (high quality images straight from the source)
Pixiv - āā (high quality images straight from the source)
Rule34 [us] - ā (NOT .xxx)
Gelbooru - āā (Large database, quick and easy to gather)
Rul34 [xxx] - (rate limits)
e621 - (rate limits)
hitomi-la - (download entire datasets)
imhentai - (download entire datasets)
pawchive - (New version of kemono)
If you plan to use my settings I would recommend gathering ATLEAST 55-60 images before proceeding to the next step however the more (quality images) the better.
**The biggest factor is dataset congruence***
Chosen images should be clean, good resolution, and look stylized
DO NOT CHOOSE IMAGES THAT ARE:
blurry
cropped
messy/cluttered (if the rest of the dataset isn't like that)
Looks like someone else could have made it
LOOKS DIFFERENT THAN THE OTHER IMAGES/STICKS OUT LIKE A SORE THUMB
āBasically don't give your model room to fail, high quality inputs = high quality outputs
Gather and save these images to a folder however you see fit
Personally, I asked Claude to build me an extension/script does this for me
Ask claude/chatgpt: Create me an extension that grabs the biggest image from each tab in a window and saves it to a folder. Basically create me an image grabber that I can use on websites like (whatever you want)
Setup/protocol
1. Open the images I want to save in a new window
2. Once I have all the images (60+ tabs) I press a button to activate my extension that automatically downloads all images to a set folder
Step 2 - Labeling/Tagging (25% of a models grade)
This section will be pretty straight-forward since a script will be doing the work for you, all you have to do is make sure it's set up correctly and verify the output.
Civitai already comes with a great built-in tagger
In the case that it's down here's the one I use:
I use joytag as the trigger model, its actually better than civitai's tagger
Here's the settings I recommend if you use Civitai's tagger:
when making a style model style do not use their regular name as a trigger word, a lot of base models have styles built-in and those can interfere with your model
To assist training the artist style you can add words like comic or flat color to the dataset where it applies
āāā Once the dataset has been labeled it is imperative that you go back and look through each image to make sure they're up to standard.
I can't tell you how much buzz I've wasted by being too lazy to stop bad data from slipping through the cracks
Step 3 - Training (35% of a model's grade)
After you've verified that your dataset is tagged and clean it's time to train!
Take a quick glance to make sure everything is correct
For my base model I always go with Illustrious XL 2.0, a lot of checkpoint models are built on it which allows my models to be more universal
Illustrious 0.1 is also good (but older)
āTraining on custom models is more expensive but yields higher quality results. If you're just learning the ropes/experimenting the built-in SDXL illustrious option could be for you
AI-Toolkit so buns, use kohya. I will not elaborate
Steps, image count, repeats, epochs, batch size
I set my batch size to 1 so the model is trained deeper image by image
I worked with Claude to develop the parameters I use now, here's an excerpt explaining what I found:
āHere's an simple explanation of the importance of step count:
This next section is pretty simple and doesn't change much across all different model types (characters, styles, concepts, etc.)
Keep tokens is the amount of trigger words that get appended to the front of the dataset instead of shuffled, if you have more than one make sure you change that value
I'll briefly explain some of these next few parameters that are most important
āIf you want to get good at making loras and troubleshooting you need to know what these affect and by how much
PLEASE CONTINUE TO EXPERIMENT, THERE'S ALWAYS ROOM FOR IMPROVEMENT / 1 MORE VERSION
If you want to blindly copy my settings you can but don't come back with questions that could be answered by understanding what the knobs you're turning do
Unet LR (Learning Rate)
This is your most important quality setting, the Unet is the part of the model responsible for visual generation. Think of it like adjusting how hard you press a pencil when tracing a drawing. Too light and the trace barely shows. Too hard and it tears through the paper.
The sweet spot I found was 0.00025 for illustrious styles
Text Encoder LR
The text encoder is what the model uses to understand language, it connects your words to visual concepts. Therefore this LR controls how strongly your trigger word and caption tags get associated with the model you're training.
I sweet spot I found was 0.00007 for illustrious styles
Min SNR Gamma
During training the model learns from images at many different noise levels from nearly clean to heavily corrupted/noisy. SNR Gamma controls how much weight is given to the harder, noisier examples versus the cleaner ones.
Lower values push the model to learn more from noisy timesteps, which helps capture fine details like linework and subtle shading. Higher values focus more on clean timesteps, producing safer but more generic results.
The sweet spot I found was 4 for illustrious styles
Network Dim and Alpha
Network Dim determines how much information the lora can store. Think of it as the size of the model's notebook. Higher dim means more capacity to capture fine details like how an artist renders eyes, hair strands, or linework edges.
Network Alpha controls the effective learning rate scaling inside the network. The standard practice is to set Alpha at half of Dim.
The tradeoff is always file size versus detail capacity. For most style loras targeting artist fidelity, 40/20 is the sweet spot (~272 MB per lora)
Noise Offset
Noise offset teaches the model to generate images with a wider dynamic range as in deeper shadows and brighter highlights. Without it, outputs can look slightly washed out or lacking contrast.
The risk is that too much noise offset over-enhances edges and fine details, producing that overly crisp or artifacted look.
ā0.02 is the sweet spot I found but it makes the model very sensitive to bad data, this is why I said to double check your inputs
Optimizers
The optimizer is the algorithm that actually applies weight updates during training. Different optimizers have different speeds, memory usage, and how they handle learning rates.
AdamW8bit is my recommended choice for style loras. it's memory efficient, well tested, and gives you precise control over your learning rate. It rewards knowing your settings but produces the most faithful results when dialed in correctly.
Other options like Prodigy adapt the learning rate automatically, which sounds appealing but tends to produce more stylized, interpretive results which is good for some use cases but not ideal when your goal is maximum fidelity to the source material.
š You did it!!!
If you followed all of these steps correctly you should have a trained model wrapped up in a safetensors file, wherever you go from there is your choice!
āI found sharing my models here to be very rewarding!
Uploading models to the site consistently gets you inducted into the early-access program where people will pay buzz to download/generate with your models
You will never pay for buzz again
Thank you for taking time to read this article, I wish you the best of luck on your model-making journey! If you do end up having questions shoot me a DM or leave a comment and others can help.
bonus photo