240_mado(style)版本Pony_v1.0 (ID: 514049)

FLUX:

sampler/scheduler : euler /beta
step : 20
cfg : 3.5
guidance : 3.5 ( raising to 5 might give better hands)

All preview images include the comfyui workflow and prompt.

Training : 64 images, captions are included in the training data (not the images) and were a mix of chatgpt4 + joytag
Auto learning rate : 1e-05 with a minimum of 1e-06
Unet learning rate : 1e-04

PONY:
/!\ WARNING /!\
The dataset used did not contain any nsfw data, to get the effect of the lora on nsfw prompt you will need to UP the weight.

Use original PonyDiffusionV6XL for maximum effect for Pony version.

Suggested settings (PONY only) :

WEIGHT:
0.8 to 1.3

GEN:

25-35steps
5-7cfg
Euler A

Highfix res:

20 steps
5-7 cfg
Euler A
0.5 denoise

+facedetailer recommended

Trigger :

style 240_mado

The lora works without any trigger but the effect will be stronger with it.

Extra trigger :

flat color, (long neck:0.35),

Negative :

It likes to make little red string on the hair on the side, you can use "red string" in negative to get rid of them.

Works best with PonyDiffusionV6XL, effect is less strong on other Pony base but still noticeable.

Sorry for the lack of metadata on some previews, the comparaison previews should have every settings, getting comfyui metadata on Civitai is a pain...

This first version lacks a bit in term of diversity, I will eventually revisit this lora to add to the original dataset to get more diversity.

This style comes from these 2 videos that I love :

描述:

训练词语: style 240_mado