V2.3 is the most stable version, please update your LoRA to the latest version.

-

If you don't mind, please help me by posting your generated images and put my LoRA as the resource so i can use it to re train and improve my LoRA since it's difficult for me to do it by myself.

-

Green: Recommended | Yellow: Alternative | Red: Worst | Orange: Im not sure / haven't tested it yet

-

My parameter:

Base Model: MajicMixRealistic v2.5 / v7

Clip Skip: 1~2 | Not sure but people says 1 is better, 2 means bad trained LoRA.

Sampling Steps: 20~40 | For me 30 is the best so far.

Sampling Method: Euler / Euler A or DPM++ SDE Karras

CFG Scale: 4~7 | I use 4 or 6(when i use Dynamic Treshold), it can be affected by steps.

Hires Upscaler: R-ESRGAN 4x / R-ESRGAN 4x+

Hires Upscale by: 2~2.5

Hires Steps: 15~25 | No fixed number here, i keep changing it and looking for best result.

Hires Denoising: 0.35~0.4

You should try your own settings looking for the best result.

-

Notes:

• V2.0 : is the stable version, but somehow it feels like i created Woman LoRA instead of a young girl LoRA.

• V2.1 : I used new method, and i will say v2.1 is a failure since we need to put 1.10~1.50 strength for the best result, but it was better in appearance than V2.0.

• V2.2 : I used new method and improved dataset. I personally liked the V2.2A face because it's really similar to the dataset, but V2.2B has better body proportion in my opinion. I can't choose one of the version so i decided to publish both version.

• V2.3 : New and better method, same dataset from the previous version. As for now this is the most stable version of Asian Beauty LoRA. Please use ADetailer if the output below your expectation.

-