ghost
Sign In

pretty Mix Girl

232
2.6k
10.2k
15
Updated: Jun 8, 2023
character
00079-3703839733-chilloutmix_NiPrunedFp32Fix.png
00093-3703839733-chilloutmix_NiPrunedFp32Fix.png
00084-3703839731-chilloutmix_NiPrunedFp32Fix.png
00092-3703839730-chilloutmix_NiPrunedFp32Fix.png
00091-3703839731-chilloutmix_NiPrunedFp32Fix.png
Images hidden due to mature content settings
Verified:
SafeTensor
Type
LoRA
Stats
2,626
10,191
Reviews
(232)
Published
Jun 6, 2023
Base Model
SD 1.5
Hash
AutoV2
EB74919742
default creator card background decoration
mixboy's Avatar
mixboy
00079-3703839733-chilloutmix_NiPrunedFp32Fix.png
00093-3703839733-chilloutmix_NiPrunedFp32Fix.png
00084-3703839731-chilloutmix_NiPrunedFp32Fix.png
00092-3703839730-chilloutmix_NiPrunedFp32Fix.png
00091-3703839731-chilloutmix_NiPrunedFp32Fix.png
Images hidden due to mature content settings

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

Pretty Mix Girl has a pretty face with a more natural look and feels for a more realistic appearance, by mixing multiple LoRAs

  • mix Twice Jihyo (0.3) + Minju Angle (0.2)+ Kim minju (0.2) for a pretty face.

    • but, mix No.183 (0.1) + No.175 (0.1) + No.177 (0.3) to wash out the original direct appearance personality.

  • add a bit of anime LoRA Ririka Momobami to add a pretty attribute.

    • a small amount of the ratio of anime LoRA makes LoRA prettier.

  • add high-quality LoRA a bit, like KoreanDollLikeness and JapaneseDollLikeness these make LoRA more robust, it has cleaner skin nice smile, etc.

    • in this case, JapaneseDollLikeness weights are higher for eyes,

  • Randomly selected LoRAs to add some properties

  • use OUTALL1 weights preset only - remove all IN* specific attributes.

    • OUTALL1:0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1 (BASE == 0) (OUTALL case, BASE == 1)

Merge Recipe

Using LoRA Block Weight extension I've tried to find a good mix ratio as follows:

<lora:Ririka Momobami:0.1:OUTALL1>

<lora:koreandolllikenessV20_v20:0.15:OUTALL1>

<lora:JapaneseDollLikeness_v15:0.3:OUTALL1>

<lora:xelor_TwiceJihyo_2.5:0.3:OUTALL1>

<lora:NO183:0.1:OUTALL1>

<lora:NO175:0.1:OUTALL1>

<lora:NO177:0.3:OUTALL1>

<lora:mjftwv2:0.2:OUTALL1>

<lora:kim-minju-23-v1:0.2:OUTALL1>

<lora:Liliya_v1.0:0.1:OUTALL1>

<lora:Chinese doll idol Round face:0.25:OUTALL1>

(in this case, OUTALL1:0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1)

  • use "merge to checkpoint" method supported by the SuperMerger extension

    • fill in the LoRA form: "Ririka "Momobami:0.1:OUTALL1,koreandolllikenessV20_v20:0.15:OUTALL1,JapaneseDollLikeness_v15:0.13:OUTALL1,xelor_TwiceJihyo_2.5:0.3:OUTALL1,NO183:0.1:OUTALL1,NO175:0.1:OUTALL1,NO177:0.3:OUTALL1,mjftwv2:0.2:OUTALL1,kim-minju-23-v1:0.2:OUTALL1,Liliya_v1.0:0.1:OUTALL1,Chinese doll idol Round face:0.25:OUTALL1"

    • 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:prettyMixGirl_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 pretty Mix Girl V1 is obtained!

only "OUTALL1" weight blocks were selected to reduce possible interference by Lora "IN*" weights. (OUTALL1 is user defined setting (used by Lora Block Weight extension ) as "OUTAL1L:0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1" that means set BASE == 0 and only OUT03-OUT11 weights are used.)

How to use

  • the recommended ratio is 0.5~0.7 like as

    • <lora:prettyMixGirl_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.

  • when using Dynamic-Thresholding or a high CFG-Scale, lowering the weight ratio will be helpful.

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

License

Like as ChillyMix, its license has no commercial restriction. it is open and free:

so It is possible to release it without any restriction!

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