New Release: 林忆宁，古户绘梨花 (010, Furuki Edoka)
Dataset:
Images chosen from 010's selfie set, specifically during her days as an acitve member of SNH48's Team X.
Training:
This LoRA model is trained with Kohya_ss GUI. Some of the training parameters are listed as following:
Base model: ChilloutMix version 6
Step:
100 steps per image, 2,000 in total(check versions)Epoch: 1
Network dim: 128
Clip skip: 2
Learning rate: 0.0001
Max resolution:
768,768(check versions)
Tips:
I suggest using this model at 0.8~0.9 weight for best results. Going above 0.9 usually causes distortion in face.
Pairing this LoRA model with prompts like close-up view or portrait is recommended. It works poorly when generating full-body pictures. (only true for v1.0)
Low CFG scale around 3.0 to 6.0 is recommended. Turn it down if you are seeing oversaturated images.
Using this model to generate NSFW content is possible, but highly dissuaded.
Reach Out:
七海电报群：https://t.me/nanamivr
宁宁肉体厨：https://t.me/nanami_ai
Any questions, suggestions and ideas on improving the model are welcomed.
Description
Retrained the model with 70 pics resized to 512×??0 as oppose to 20pics resized to 512×512 in v1.0. 30 steps each image which makes 2,100 steps in total.
v2.0 can do a better job of handling pictures with upper body, while v1.0 can only generate portraits properly.
v2.0 is less 010-like, in the sense that the loss function is greater at the end of the training as to v1.0.
FAQ
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.