Months ago someone has trained one, but that has gone so I trained it again by myself.
Also I finally figure out what data cause my models not good enough, so I'm gonna doing some updates on existing model. Maybe will write my experience of LoRA training by article.
How to use:
Trained Resolution 768x768
Weight 0.6-0.8
Base for costume:
Colorado, blonde, braid, blue eyes,
Original Outfit:
ColoradoOG, blonde, braid, garrison cap, headgear, necktie, backless outfits, grey dress, elbow gloves, pantyhose,
Description
FAQ
Comments (2)
Thanks for the cute BB.
I'm always interested to hear what methods work for other people in LoRA development!
Thanks. In conclusion is, all those updates for optimize LoRA training (except multires noise), are have drawbacks that is not suggested to use.
SNR noise will cause the model became easily overfitted, noise offset and LoRA block training will cause lack of details, and the gradient checkpointing is the biggest problem - it do save some VRAM, but it is the cause of concept melting, like my Kashima model had.
The best training of LoRA guide for now, is not to use any new gimmicks since March 2023. Except multires, that's a good feature and really useful.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.