Sure I can tell you my process, to be honest I kinda winged it from the start and it managed to turn out really good so I kept using more or less the same process, it’s a little long so maybe copy paste it to save it somewhere but I’ll run you through my entire process if it helps:

Even though I now generate my images locally I still make loras through civit as I find their tools easier to use.

I use Wai NSFW Illustrious v14

My data set image counts vary a lot but the main things I noticed are:

45-50 is usually my minimum if the character has very little actual content and usually makes a solid model

30-35ish is my hard minimum and not recommended as it can still work but the model is noticeably less flexible and will have some issues

If possible I always aim around 90-110 for the best balance of quality and buzz economy, the highest I’ve ever done is like 190 and I didn’t really notice that much of an improvement

I usually collect the datasets by screenshotting from my phone as I flip through the pages online, I then transfer them to my pc using OneDrive, then I use an image converter (I like irfanview 64) to convert the images to jpg, then I use an app called pixel pruner (on GitHub) that lets me crop the images faster, when they crop they get converted to pngs so I run the cropped images through the image converter again so they are jpgs (from my experience jpgs transfer much faster onto civit and the auto tagged has no issues and tags faster), something I’ve noticed is that the resolution of the original data set is not that important as hi res fix generation fixes any blurring.

When I use the auto tag, in the prepend just put the characters most important permanent features that they always have e.g hair colour, eyes, body type, hairstyle, in the blacklist I usually put the generic blacklist tags that most people use e.g bad quality, mutation, deformed. For the append tag I usually just put “manga” or “manhwa” depending on what the character is from, this basically makes it so the manga tag can be swapped in and out of the positive prompt when generating whether people want the manga panel styling or just the character with no panels or text boxes, I usually put the max tags it generates up to 30 (not actually sure if it makes a difference, but 30 is what I’ve always done), I put min threshold on 0.4.

In the trigger words box above the images, I usually just do the characters name, later when preparing to post the finished model I usually add triggers groups including her name and her features that were mentioned in the prepend, also if the character wears a specific outfit I use the generator to pinpoint the perfect tags to recreate it to then copy paste those into the trigger words section

Something I started doing recently to be able to generate certain characters in monochrome and colour is to try to find some official colour pages or coloured cover art, what I noticed is even if you have a very large amount of monochrome images in your data set (>100) about 10-15 colour images will be enough for the trainer to learn it, just make sure to put monochrome in the negative prompts when generating with it as sometimes it will make monochrome/colour hybrids.

In the final review and submit page after the auto tagging is done, I select sdxl illustrious in the base model options.

The sample training image prompts don’t need to be perfect as they don’t affect the final result but they help in displaying the process as the model trains.

These are my training parameters, again I’m not sure if these are actually optimal but they are what I’ve used from the start:

Epochs: 30

Num Repeats: depends (explained below)

Train batch size: 1

Enable bucket: on

Shuffle tags: on

Keep tokens: 1

Clip Skip: 2

Unet LR: 0.00003

Text Encoder LR: 0.00003

Lr scheduler cycles: leave as it is

Min SNR Gamma: leave as it is

Network dim: 32

Network alpha: 32

Noise offset: leave as it is

Optimiser: AdamW8Bit

So I like high step counts for my models as I feel it is more accurate to the source material, the num repeats should be adjusted so that the greyed out step count is anywhere between 4000-5500, I have tried some models with over 9000 steps but I found that it made no real quality difference and just cost way more buzz

Don’t bother with high priority as I haven’t found that it makes any difference.

Usually a model in that step count sets me back about 1.8k buzz and takes ~4 hours to train

Hope this helps and sorry for such a long message 😂

@zetsubou yeah I’m probably gonna tidy up the guide a bit and post it as an article, I don’t really mind people knowing since I mainly do it for fun

@zetsubou yeah trust me, when you make your first good working Lora it gets addicting haha, you realise that as long as you have enough images of a character you can pretty much recreate everyone

@zetsubou yeah I've been planning to do that once i have time, I've seen other creators do it too