Type | |
Stats | 3,354 21,650 |
Reviews | (414) |
Published | May 19, 2023 |
Base Model | |
Hash | AutoV2 2F66BF5863 |
 x alpha_0) x (1 - alpha_1) + beautifulRealistic_brav3 x alpha_1) x.jpeg)
This is a free/open available Mixed Lora.
If you like my work, feel free to buy me a coffee at ko-fi. https://ko-fi.com/mixboy
The following LoRAs are used. All credits go to the original authors. Thanks!
LoRAs have been merged with the SuperMerger extension
Design points
Cutie Mix Girl has a realistic cute face without being too over-the-top.
downgrade the cuteness of No.177 Girl by merging No.169 Girl
No.169 Girl is boyish
add a bit more natural style LoRA like Dr.She - Dr.She is a nice example. it is not so pretty and not so cute but well made.
add high-quality LoRA a bit like as KoreanDollLikeness. it makes LoRA more robust, it has cleaner skin nice smile, etc.
you may need to use OUTALL weights preset only - so many low-quality LoRAs have not optimal IN* weights and it makes trouble. simply use only OUTALL weights to solve this issue.
Merge Recipe
Using LoRA Block Weight extension I've tried to find a good mix ratio as follows:
the obtained and used ratio is No.177 : No.169 : Dr.She : koreandolllikenessV20_v20 = 0.5 : 0.4 : 0.15 : 0.02
No.177 and No.169 are remade its dimension to 128. (the remade dimension version is named NO177, NO169 in each case.)
use only OUTALL weights like as
<lora:NO177:0.5:OUTALL><lora:Dr.She:0.15:OUTALL><lora:NO169:0.4:OUTALL><lora:koreandolllikenessV20_v20:0.02:OUTALL>
In this case, It is hard to use merge several LoRAs in 1-step and reproduce the same result. so instead, I use "merge to checkpoint" method supported by the SuperMerger extension
fill in the LoRA form: "NO177:0.5:OUTALL,Dr.She:0.15:OUTALL,NO169:0.4:OUTALL,koreandolllikenessV20_v20:0.02:OUTALL"
output file name as "MERGE1"
select checkpoint-A "v1-5-emaonly"
then press "merge to checkpoint" button and you can get the "MERGE1".
select checkpoint-B "MERGE1", output file name as "new-lora"
then press "Make LoRA (alpha * A - beta * B)" button and you can get "new-lora" (alpha == beta == 1)
to make <lora:cutieMixGirl_V1:0.6> is the default, scale the LoRA by 1.7(1/0.6=1.666...) with the "same strength" option.
now the cutie Mix Girl V1 obtained!
In this case, only "OUTALL" weight blocks were selected to reduce possible interference by Lora "IN*" weights. (OUTALL is defined by Lora Block Weight extension as "OUTALL:1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1" that means BASE,OUT03-OUT11 weights are used.)
How to use
recommended ratio is 0.5~0.7 like as
<lora:SweetyMixGirl_V1:0.6> (0.6 is designed ratio)
lower ratio weights could be a good starting point for non Lora-friendly-Models.
the higher ratio could result in overfitted behavior. in this case, lower your ratio.
What's the difference
and what are the advantages of this unlike anything other LoRA girls?
it is not a trained LoRA, it is a mixed LoRA.
its recipe is open, anyone who wants to change the mix ratio and wants to get proper results can change their recipe based on it.
works nicely with the ChilloutMix, and also works fine with ChillyMix or Muse_v1 or other LoRA friendly models.
Recommends
default Clip skip: 2, ETA noise seed delta: 31337
for BRA v5 case, Clip skip:1 is preferred.
Useful AUTOMATIC1111's webui extensions
webui model toolkit to examine Loras and models
civitai extension to add several useful feature for civitai
License
Like as ChillyMix, its license has no commercial restriction. it is open and free:
Dr.She and KoreanDollLikenessV20 could be merged and the merges could "Have different permissions when sharing merges" license.
No.177 Girl , No.169 are free! thanks again!
CreativeML-Open Rail-M: This model permits users to:
✔Use the model without crediting the creator
✔Sell images they generate
✔Run on services that generate images for money
✔Share merges using this model
✔Sell this model or merges using this model
✔Have different permissions when sharing merges
Support me
If you like my work, feel free to buy me a coffee at ko-fi. https://ko-fi.com/mixboy