Thoughts on LoRA Training Pt 2: Where to Train
This is a pretty quick follow up, but there were some immediate "where do I start" questions I want to answer.
First and foremost, if you have never trained a LoRA before, start somewhere that has presets - as in a notebook or platform that has preset trianing parameter values. I beg of you.
The reason I insist on that is - if you go in and adjust settings before you know if you understand how to curate a dataset and captions, then you will never known if the issue is your data or your parameters. It will be a nightmare to diagnose and very discouraging. Also, most of these platforms or default settings aren't arbitrary - they've worked for someone at some point, so presuably you should be able to get them to work!
So without further hesitation, here are a list of tools I've used and some I haven't but I've seen good results from.
OPEN SOURCE
I think open source can be the way to go - it is certainly how I started - and it just depends on your willingness to sacrifice time for utility in many cases.
The simplest and TLB is fantastic. Really, if you want to just not overthink things and get started, use his runpod. I recommend keeping your captions super short with a unique token + a couple words description format. If that, honestly, Ben has always recommended just to use the unique token. Either works.
The only challenge of just using the unique token is that you will need to ensure you have a balanced dataset. As in - enough images but not too many (10-30), no doubles, consistency of the concepts you want to keep and variability of the things you don't want the model to learn.
Kohya is also great - I just hesitate to put it in front of new people because it allows you to adjust so much. I feel like that can be a recipe for disaster. But, ultimately, a gold standard.
PAID SERVICES
Truth be told, I haven't met a silver bullet platform, but here are all the ones I am aware of with notes on whether I've used them.
Scenario - I actually helped develop the training presets at Scenario, so I quite like it. I believe it is still unlimited trainings for account holders, however there isn't a way to export your models which I find really challenging. I usually just use the presets.
CivitAI - I don't love the vibe of Civit, but I would be misleading you if I said their training doesn't work. I even trust their crazy autocaptions. Typically I do 20 epochs with 17 repeats and Clip Skip of 2.
Pimento - I am pretty impressed with the presets they have for illustration on Pimento. I haven't dug into it too deep, but ultimately I think it has a lot of promise as a service.
Leonardo - I haven't spent a ton of time using Leonardo for training, but I wasn't wholey disappointed when I did. I do think their finetuning team makes amazing things and would love to see their training updated more to reflect that.
Layer - I haven't used their service much myself, but I have heard people get really strong results.
EverArt - For people who want a really really simple setup without too many changes, I think EverArt shows a lot of promise.
Astria - I actually trained my first time on Astria in 2022, and I think they have the consistent character/object set up really well established. I haven't revisited it yet but it is on my list.
So - hopefully this list helps you!
