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
-
don't try to make it too beautiful/cute/sexy
-
No.177 Girl is too cute but has a nice style.
-
-
preparing a 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 LoRA have not optimal IN* weights and it makes trouble. simply use only OUTALL weights to solve this issue.
Merge Recipe (V1)
Using LoRA Block Weight extension I've tried to find a good mix ratio as follows:
-
the obtained and used ratio is Dr.She : koreandolllikenessV20_v20 : No.177 = 0.2 : 0.1 : 0.6
-
use only OUTALL weights like as "<lora:Dr.She:0.2:OUTALL><lora:koreandolllikenessV20_v20:0.1:OUTALL><lora:sd-No.177-000014:0.6:OUTALL>"
-
in this case, No.177 Girl is 32 dim, but SuperMerger can "remake dimension" to 128. (the remaked dimension version is named NO177 in this case.)
-
merge with the SuperMerger extension "Dr.She:0.4:OUTALL,koreandolllikenessV20_v20:0.2:OUTALL,NO177:1.2:OUTALL" (x2 multiplied and check "same strength" option.)
-
it's not easy to reproduce the procedure but in this case "same strength" is used.
-
-
then merge and obtain the Sweety Mix Girl V1
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.)
Merge Recipe (V2)
Use merge to checkpoint menu instead of Merge LoRA method.
-
same as V1: Dr.She : koreandolllikenessV20_v20 : No.177 = 0.2 : 0.1 : 0.6
-
same as V1: use only OUTALL weights like as "<lora:Dr.She:0.2:OUTALL><lora:koreandolllikenessV20_v20:0.1:OUTALL><lora:sd-No.177-000014:0.6:OUTALL>"
-
same as V1: in this case, No.177 Girl is 32 dim, but SuperMerger can "remake dimension" to 128. (the remaked dimension version is named NO177 in this case.)
-
V2: use merge to checkpoint menu with the SuperMerger extension
-
fill in the LoRA form: "Dr.She:0.2:OUTALL,koreandolllikenessV20_v20:0.1:OUTALL,NO177:0.6: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:sweetyMixGirl_V2:0.6> 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 V2 obtained!
※ "merge to checkpoint + "Make LoRA (alpha A - beta B)" method is a much more reproducible method than the"make LoRA" menu.
How to use
-
recommended ratio is 0.5~0.7 like as
-
<lora:SweetyMixGirl_V1:0.5> or <lora:SweetyMixGirl_V1:0.6>
-
-
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
What are the advantages of this unlike anything other loRA girls?
-
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, also works fine with ChillyMix or Muse_v1
Recommends
-
default Clip skip is 2, ETA noise seed delta: 31337
-
for BRA v5 model, 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.177 Girl is free!
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
描述:
2023-06-07 - updated version released
训练词语:
名称: sweetyMixGirl_V2.safetensors
大小 (KB): 147567
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
