Type | |
Stats | 1,511 955 |
Reviews | (220) |
Published | May 22, 2023 |
Base Model | |
Hash | AutoV2 7CC864ABA0 |
 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
Honey Mix Girl has a face with a more natural look and less makeup, for a more realistic appearance.
mix No.175 Girl and No.179 Girl
both are unique and sexy
mix No.160 Girl LoRA background brightness
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.
if you want to add more makeup and cleaner skin, add "<lora:koreanDollLikenessV20:0.2>" line at the end of prompt.
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:
No.175 : No.179 : No.160 : Dr.She : koreandolllikenessV20 = 0.3 : 0.4 : 0.15 : 0.1: 0.01
No.175 , No.179 and No.160 are remade its dimension to 128. (the remade dimension version is named NO175, NO179, NO160 in each case.)
use only OUTALL weights like as
<lora:NO175:0.3:OUTALL><lora:NO179:0.4:OUTALL><lora:NO175:0.15:OUTALL><lora:Dr.She:0.1:OUTALL><lora:koreandolllikenessV20:0.01:OUTALL>use "merge to checkpoint" method supported by the SuperMerger extension
fill in the LoRA form: "Dr.She:0.1:OUTALL,NO160:0.15:OUTALL,NO175:0.3:OUTALL,NO179:0.4:OUTALL,koreandolllikenessV20_v20:0.01: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:honeyMixGirl_V1:0.6>" is the default weight, scale the LoRA by "1.66666666666666666"(1/0.6=1.666...) with the "same strength" option.
now the honey 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
the recommended ratio is 0.5~0.7 like as
<lora:honeyMixGirl_V1:0.6> (0.6 is the designed ratio or below 0.6)
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.
this work is done 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 features 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.175 Girl , No.179 Girl , No.160 Girl 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