LoRA użytkownika

SD

Ky. Style

Mio with expression sulk

Okładka usunięta: Naruszenie zasad treści

Nie przeszło naszego sprawdzenia treści — bez obaw, Twój model nadal jest w pełni przeszukiwalny i użyteczny! Po prostu zamień okładkę na taką, która będzie odpowiednia dla wszystkich.

Przesłano
1 paź 2023, 10:49
Użycia
725
Opinie
Brak recenzji
Słowa kluczowe

KyStyle,

Uprawnienia
  • Allow image generation & sharing
  • Zezwól użytkownikom pobierać Twój model
  • Zastosowania komercyjne

Opis

https://civitai.com/models/153389/ky-style-hentai-nsfw

Training: STEPS: 30 EPOCHS: 10 CLIP SKIP: 2

Trigger Words: KyStyle

Hello! I did my second style lora - ky. style (previous was Xtermination style).

0.9 - 1 strength of Lora is the best radius (If you combine with character LoRA I recommend to decrease weight of character LoRA)

Clip skip 2

Some necessary tokens for better experience:

KyStyle, 1girl

Important tips:

There were too many arts with Minamoto Raikou in dataset. So if you feel that LoRA tries to generate her and you don't want it put "minamoto no raikou (fate)," in negative prompt.

Put in negative prompt "empty eyes" if SD generates eyes without irises.

Most of the artist's pictures were censored. So if you do NSFW scenes it is good idea to put "censored, bar censor, blank censor" in negative prompt.

Use vanilla anime models without high realistic vibes for saving artist's style. I recommend you anything v5 and divineanime. Also I suppose it will be work well on bloodorangemix.

If you feel that arts are too colorful you can change kl-f8 anime vae to animevae (orange.mix vae / novelAI vae). Also you can try blessed vae.

Main advices for all LoRAs:

I recommend you next prompt order for NAI-based mixes (you can see it in my preview arts):

(highres, high quality:1.2) ("masterpiece" and "8K" kill some concepts. Be careful if you try some complex clothes or concepts.)

Also put lighting / shadows, intricate details if you need some extra effects. Sometimes it doesn't work if style of anime or game was baked in LoRA or anime checkpoint mixes lost some concepts which exists in vanilla SD 1.5.

Activation word (from lora civitai page, usually character name). Describe with danbooru tags character activity, expressions, camera position;

Hair details, face details, jewelry, crowns, hats, etc.

Clothes and shoes; Or nude.

Body details (large breasts, cleavage). Genitals if you want to generate it.

Background.

Advices for upscaling / fixing:

use highres.fix (If you are sure that your prompt is easy for SD); I prefer x4_Fatal_anime_500000_G, animesharp, foolhardy_remacri, siax with 1.5x - 2x upscale. 0.36 denoising and 15-18 Hires fix is enough.

Do not use "latent" upscaler because it works worse with denoisning < 0.5 but you don't want so high denoising stength because it changes your image instead of adding details.

use extention after detailer for better faces. It is better than inpaint.

use SD Upscale / Ultimate SD Upscale script in img2img after highres fix if you want 4K or more resolution with adding details (don't use Extras for it).

Use inpaint and sketch for other details;

Use Photoshop / Photopea extension for fixing hands, textures and other.

I will be glad to receive feedback from you and a demonstration of your art with my LoRA!

Komentarze

Stworzono przez Ky. Style