MM_Girl_ Super realistic high-quality Asian beauty

Introduction: Using my own sister as a test subject, training the realistic and touching Lora model based on ChilloutMix. The base model Lora is trained on the ChilloutMix large model and enhanced with majicMAX, excelling in realistic and lifelike portrayal in photography genres.

Features: Excellent facial generalization, super natural human-like portraits, enhanced facial details, skin texture, and texture.

Idea Overview: As of now, AI-generated portraits have become too industrialized, dull, lacking uniqueness, and suffering from a serious lack of distinctiveness. In order to reject the stereotypical AI faces and the unremarkable faces of internet celebrities, this model was trained to add uniqueness and beauty to AI-generated portraits.Each has its own beauty.

Moreover, perfection doesn't exist, and in my understanding of realism, a mole, some freckles, some highland redness, or rough skin might also be a form of beauty, isn't it?

Recommendation for Generation: Pairing with a large realistic model such as the majicMAX series, the Ink Ghost Humanoid series, XXMIX series, ChilloutMix or other realistic large models.

Recommended Keywords:ray tracing, photorealistic, Detailed beautiful delicate face, finely detailed beautiful eyes and detailed face, Ray tracing, Cinematic Light, light source contrast to enhance facial details and lighting effects.

Weight recommendation: (0.5-0.9)

Recommended Steps: 35-50

VAE: vae-ft-mse-84000-ema

Recommended enabling high-definition restoration.

High-definition restoration settings:

Upscaling Algorithm: 4X-UltraSharp   Redrawing Magnitude: 0.2-0.35     HD Repair Sampling 25

Recommend using the ADetailer facial enhancement plugin   place LORA in ADetailer's prompt,

As well as the CDTuner plugin

Plugin Syntax: Input directly, or adjust by dragging within the plugin

<cdt:d1=2;col1=-3>

<cdt:d2=2;hrs=1>

<cdt:1>

<cdt:0;0;0;-2.3;0,2>

<cdt:0;0;0;-2.3;0;2;0;0;1>

It’s good to listen to advice and have a full stomach. If you have any thoughts or suggestions, you can reach me atWX: BLUE601159

Comparing V2.0 version with V1.0 for this model, adding some data visualization, enhancing training, good generalization, and improved skin texture and complexion.

(P.S.: Alchemy is mystical; to get it right, one has to endure many setbacks. The person involved is not very satisfied, but soldiers on nonetheless.)

Model fitting and generalization test image:

Lastly, a side note to recommend a sampler that I frequently use – DPM++ 2M alt Karras

To add, open with Notepad stable-diffusion-webui/modules/sd_samplers_kdiffusion.py  Add at line 38:

from tqdm.auto import trange

@torch.no_grad()

def sample_dpmpp_2m_alt(model, x, sigmas, extra_args=None, callback=None, disable=None):

   """DPM-Solver++(2M)."""

   extra_args = {} if extra_args is None else extra_args

   s_in = x.new_ones([x.shape[0]])

   sigma_fn = lambda t: t.neg().exp()

   for i in trange(len(sigmas) - 1, disable=disable):

       denoised = model(x, sigmas[i] * s_in, **extra_args)

   return x

k_diffusion.sampling.sample_dpmpp_2m_alt = sample_dpmpp_2m_alt

Simply save after completion