My first attempt at fine-tuning an illustrious model using a LoKr. The model has been trained on a set of 6000 high quality images that were captioned with JoyCaption.
The LoKr is still a bit under trained and I recommend using high strengths, but you might want to test that yourself. The base model used to train the LoKr is https://civarchive.com/models/1323352/illustrious-ultrareal-6gb which seems to be one of the best photoreal illustrious chekpoints out there as of today.
The basic idea is to use a large LoKr to finetune a model, I originally wanted a larger LoKr, but I messed up the setup and ended up with a smaller set of parameters. In the end that might have been a good thing.
All in all, I am happy with the result, but I do not consider it a perfect product yet. It could have been done better.
Description
First try
Comments (6)
@generatorofthingsreal
I tried this LyCORIS with the model you use as a sample.But nothing changed. I posted a comparison image in the this page Gallery.
I tried it with other Illustrious models and got the same results. Do you know the cause? I'm using version 1.9.4 of the A1111.
I have no idea. I would guess user error, that you are doing something wrong. That is at least the cause of 99% of my problems, when something does not work then I have done something wrong. Are you sure you are doing it correctly?
If its not user error, then I have no idea. It is a bit under trained and need a high strength, but you should see something at 1.
@generatorofthingsreal Yes, it's possible that my environment is the cause. However, the rlsk_pony and rlsk_sdxl you created change properly. This is the first time I've encountered a case like this where there is no change. Was the learning method of NIRL a special method different from rlsk_pony and rlsk_sdxl?
@fantaseed yes, but that should not matter.
@fantaseed I just did a test and posted the results, works well for me.
@generatorofthingsreal It's just a guess, but I think I've found the cause.
Two months ago, I wrote the following about an error where illustrious ULtraReal couldn't be loaded on the A1111.
https://civitai.com/models/1323352/illustrious-ultrareal-6gb?dialog=commentThread&commentId=725473
Probably, illustrious ULtraReal can't be loaded on my A1111 environment, so it's possible that NIRL trained on that model isn't effective either.
I can load illustrious ULtraReal on comfi. So, as expected, NIRL changed properly on comfi.
It's unclear whether this applies to all A1111 users.