About
I've noticed that there wasn't a LoRA of G41 from GFL and decided to give it a go. Recommended to use at weights 0.65-0.8.
All of the preview images were generated with the same prompts and nothing was done to them after the initial generation. I did leave out any inconsistent results and generally the results that didn't look amazing or close to the character, so keep that in mind, but it's fairly consistent if you use more words like: blonde, heterochromia, etc. in your prompt.
The LoRA was trained on Anything V5 and the previews were also generated using it, I've not tested it on any other models but I'm sure it will work fine on others.
Update!
Fixed several issues that were pointed out. The model should work much better now than before. I still recommend adding in prompts like: heterochromia, blonde hair, etc. but just using the g41 trigger word should get the general character concept out now.
The LoRA should also work much better on other models, I only tested 5 so there could be some that don't take it well but hopefully now it should be good.
Thanks to yakumonagi246 for pointing out the problems in the training for the LoRA, I probably would never have noticed and would've kept training them wrong otherwise.
Description
Fixed captioning problem and trained on NAI model instead.
Should work much better with other models and not have as many issues. Generally retains a specific look and outfit (heterochromia is a bit inconsistent, much better results specifying eye colours) when using only the "g41" trigger word but using other words to define the outfit or general appearance should work much better in this version.
Thanks to yakumonagi246 for pointing out these errors in the LoRA, hopefully this version is much better trained.
FAQ
Comments (2)
G41 is so cute, even though I dont play this game
Out of curiosity, what 'mistakes' were you making when training the LoRA? Not pruning tags like heterochromia and blonde hair, so they were required to generate properly?
Sometimes I greatly prefer models that require you to tag things like this, so you can be more flexible with changing the character's feature or clothes. It's definitely a trade-off though.
The main mistake was the configuration of kohya, I left out a field which set it to try and find .caption files instead of .txt. This meant it was only reading a single tag instead of the ones I had set. This caused the LoRA to consider everything in the dataset regardless of what it was to be g41. I think that's what happened to the best of my knowledge.
I also agree, I like certain parts of the character to be flexible like clothes but key features like hair or eye colour to appear on a single trigger word (even though the AI sometimes just refuses), but from my knowledge I think you prune tags that you want the AI to associate with a trigger. I could be wrong though.