dolly Mix Girl

This is an attempt to make a free/open available 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

• Dr.She v1.0 Lora by @stablecolab2023

• KoreanDollLikeness (v2.0) by @Kbrloras

Merge Recipe (V1)

Using Lora Block Weight extension I've tried to find good mix ratio like as follow:

• obtained and used ratio is Dr.She : koreandolllikenessV20_v20 = 0.4 : 0.2

• use only OUTALL weights like as "<lora:Dr.She:0.4:OUTALL><lora:koreandolllikenessV20_v20:0.2:OUTALL>"

• merge with the SuperMerger extension "Dr.She:0.8:OUTALL,koreandolllikenessV20_v20:0.4:OUTALL" (x2 multiplied and check "same strength" option.)

• then merge and obtain the Dolly 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 = 0.4 : 0.2

• same as V1: use only OUTALL weights like as "<lora:Dr.She:0.4:OUTALL><lora:koreandolllikenessV20_v20:0.2: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:0.4: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:dollyMixGirl_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 dolly 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 0.4~0.7 like as

  • <lora:DollyMixGirl_V1:0.5> or <lora:DollyMixGirl_V1:0.6>

• lower ratio weights could be good starting point for non Lora-friendly-Models

  • for BRA, with "Korean|Japanese woman" works fine

• higher ratio could results 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 results a bit can change their recipe based on it.

• works nicely with the ChilloutMix, also works fine with ChilyMix or Muse_v1

Recommends

• You can use ADetailer or DDetailer to enhance face details.

• default Clip skip is 2, ETA noise seed delta: 31337

  • for BRA v5 model, Clip skip:1 is preferred.

Useful AUTOMATIC1111's webui extensions

• SuperMerger extension

• Lora Block Weight extension

• 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.

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