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

• No.182 Girl by @dawn66666666

• No.170 Girl by @dawn66666666

• Dr.She v1.0 by @stablecolab2023

• TaiwanDollLikeness (v2.0) by @Kbrloras

• KoreanDollLikeness (v2.0) by @Kbrloras

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

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

• 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)

  • use ADetailer rather than DDetailer

• try to add "<lora:epi_noiseoffset2:1.8>" LoRA and adjust the weight ratio of 1.8 if necessary

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.

• 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