Type | |
Stats | 1,968 6,473 |
Reviews | (232) |
Published | May 27, 2023 |
Base Model | |
Hash | AutoV2 AA5CD03776 |
 x alpha_0) x (1 - alpha_1) + beautifulRealistic_brav3 x alpha_1) x.jpeg)
This is a free/open available Mixed LoRA.
The following LoRAs are used. All credits go to the original authors. Thanks!
LoRAs have been merged with the SuperMerger extension
Design points
Lovely Mix Girl has a pretty face with a more natural look and less makeup, for a more realistic appearance.
mix No.182 Girl and No.170 Girl
No.182 for cute and a smaller nose, No.170 for face style
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 KoreanDollLikeness and TaiwanDollLikeness these make LoRA more robust, it has cleaner skin nice smile, etc.
in this LoRA, KoreanDollLikeness + TaiwanDollLikeness weights are higher for a prettier face.
use OUTALL weights preset only - remove all IN* specific attributes.
Merge Recipe
Using LoRA Block Weight extension I've tried to find a good mix ratio as follows:
No.182 : No.170 : Dr.She : taiwanDollLikeness : koreandolllikenessV20 = 0.4 : 0.2 : 0.25 : 0.3: 0.3
No.182, No.170 are remade its dimension to 128. (the remade dimension version is named NO182, NO170 in each case.)
use only OUTALL weights like as
<lora:NO182:0.4:OUTALL><lora:NO170:0.2:OUTALL><lora:Dr.She:0.25:OUTALL><lora:taiwandolllikenessV20_v20:0.3:OUTALL>
<lora:koreandolllikenessV20_v20:0.3:OUTALL>
use "merge to checkpoint" method supported by the SuperMerger extension
fill in the LoRA form: "NO182:0.4:OUTALL,NO170:0.2:OUTALL,Dr.She:0.25:OUTALL,taiwandolllikenessV20_v20:0.3:OUTALL,koreandolllikenessV20_v20:0.2: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-A "MERGE1", and checkpoint-B "v1-5-emaonly", 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:lovelyMixGirl_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 lovely 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:lovelyMixGirl_V1:0.6> (0.6 is the designed weight 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: (WARNING: experimental instructions)
Clip skip:1 is preferred.
try to change your sampling method. for example, DPM2 or DPM2 Karras
or lower CFG Scale ~4
or Enable Dynamic Thresholding with the following advanced options:
Set CFG Scale 7~13 and mimic CFG Scale to ~4
Top percentile of latents to clamp: ~100
Mimic Scale Scheduler: Linear Down or what you want
Minimum Mimic Scale Schedular: ~4
CFG Scale Scheduler: Linear Down or what you want
Minimum CFG Scale Scheduler: ~4
try to change your sampling method. (e.g.) DPM++2M Karras)
try to add "<lora:epi_noiseoffset2:1.8>" LoRA and adjust the weight ratio of 1.8 if necessary
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.170 Girl, No.182 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
Donation
Feel free to donate to me if you want to support my work! at ko-fi https://ko-fi.com/mixboy