Chanson Style Adder (CSA) & Chanson Detail Enhance

LORA

An experimental style LoRAs - CSA & CDE [Disclaimer]

These LoRAs are not made for SD beginners because the CSA & CDE needs to be fine-tuned manually during the generation process

PLEASE DONT NOT USE FOR COMMERCIAL ACTIVITIES

Background

CSA_Alpha (CSAA) was trained a month ago and it worked with other LoRA well, but CSAA will not open to public

Chanson Style Adder (CSA) aims to apply Kancolle Artist style and improve the generation of Kancolle Character in terms of art style

Chanson Style Adder - Mid-Life Update (CSA-MLU) is an extra resource pack

Chanson Detail Enhancer (CDE) aims to enrich the image's detail which slove the disavandtage of CSA, lack of detail but concept art

*In other word, CDE is the offspring of CSAA

Purpose

[CSA] -Applying kancolle artists style to your generation -Reducing the amount of prompt without losing details

[CSA-MLU] -Applying more kancolle artists style

[CDE] -Assit CSA to become a compheresive style LoRA -Output more details on [Drafting State], 512*512 res generation

Info

[CSA-v0.1] -Training with 672 images, with 6800 steps -512*512 training envirnoment

[CSA-v0.2] -Training with 672 images, with 9200 steps -768*768 training envirnoment

[CSA-MLU] -Training with 318 images, with 10040 steps -768*768 training envirnoment

[CDE] -Training with 617 images, with 10020 steps -768*768 training envirnoment

Recommended set up

These LoRAs weight: 0.5 for [general case], please go to [LoRA weighting] section for specific setup

Basic prompts

[CSA]

-KonishiCSA

-KujouCSA

-ShibafuCSA

[CSA-MLU]

-HoshiakariCSA

-ShizumaCSA

[CDE]

-AlphaCDE

-BetaCDE

-GammaCDE

---BREAK LINE---

[CSA-v0.1]

-headgear, hair flaps !NEG!

[CSA-v0.2] -machinery, weapon, rigging !NEG!

[CDE] -nsfw !NEG!

Testing environment

(masterpiece, best quality), [CSAs/CDEs], BREAK, 1girl, solo, BREAK, (Character LoRA prompts), BREAK, oblique portrait, upper body shot, gradient background, white background, BREAK, <lora:CSA/CDE:?>, <lora:Character:1>

(worst quality, low quality)

Samper: DPM++ 2M Karras

Sampling steps: 30 / [ 40 (hires-fix) ]

Res: 512*512

CFG Scale: 7

Clip Skip: 2

Hires-fix: 512*768

Hires-fix steps: 30

Hires-fix Denoising Strength: 0.65

Advance Setting -LoRA weighting

[CSA-v0.1]

[0.50 - 0.60] : recommendation

[CSA: Default]

[0.00 - 0.30] : Weak

[0.31 - 0.60] : Safe

[0.61 - 1.00] : Strong

[CSA: KonishiCSA]

[0.00 - 0.40] : Weak

[0.41 - 0.65] : Safe

[0.66 - 1.00] : Strong

[CSA: KujouCSA]

[0.00 - 0.25] : Weak

[0.26 - 0.69] : Safe

[0.70 - 1.00] : Strong

[CSA: ShibafuCSA]

[0.00 - 0.24] : Weak

[0.25 - 0.65] : Safe

[0.66 - 1.00] : Strong

[CSA-MLU: Default]

[0.00 - 0.30] : Weak

[0.31 - 0.70] : Safe

[0.71 - 1.00] : Strong

[CSA-MLU: ShizumaCSA]

[0.00 - 0.24] : Weak

[0.25 - 0.60] : Safe

[0.61 - 1.00] : Strong

[CSA-MLU: HoshiakariCSA]

[0.00 - 0.35] : Weak

[0.36 - 0.80] : Safe

[0.81 - 1.00] : Strong

[CDE: Default]

[0.00 - 0.40] : Weak

[0.41 - 0.70] : Safe

[0.71 - 1.00] : Strong

[CDE: AlphaCDE]

[0.00 - 0.40] : Weak

[0.41 - 0.60] : Safe

[0.61 - 1.00] : Strong

[CDE: BetaCDE]

[0.00 - 0.15] : Weak

[0.16 - 0.69] : Safe

[0.70 - 1.00] : Strong

[CDE: GammaCDE]

[0.00 - 0.24] : Weak

[0.25 - 0.60] : Safe

[0.61 - 1.00] : Strong

-LoRA Block Weight (Extension Need to be installed)

[CSA]

MIDD: PASS

OUTD: PASS

Other: Unable to determine

[CSA-MLU]

MIDD: PASS

OUTD: PASS

Other: Unable to determine

[CDE]

MIDD: PASS

OUTD: PASS

Others: Unable to determine

-CFG Scale

7

-Compatibility

CSA + CDE + ChansonMix + Chanson's LoRAs : Highest

CSA + CDE + CetusMix/Hassaku + Chanson's LoRAs : High

CSA + CDE + CetusMix/Hassaku + others's LoRAs : Close to High CSA + CDE + other's checkpoint + other's LoRAs : Unable to determine (Need samples)

-Multi-apply weight

[CSA]

Recommended to use | for Mixing art style E.g. (KonishiCSA:0.7)|(KujouCSA:0.3)

[CDE]

?

Anyway, have fun with CSA & CDE

版本详情

基础模型 1.5
10020
10

项目权限

模型转载自: https://civitai.com/models/120602?modelVersionId=131170

转载模型仅用作交流、学习使用,不可商用。 原作者可加入交流群联系吐司工作人员认领模型。

相关帖子