bother to read descriptions

@P317cm You are just awesome dude, the images turn out fantastic

Playing around with the model, also found that you need to set 'Clip Skip' to '2'. And thanks for the VAE OP

Also, got closer to replicate OP's results by enabling "Use old karras scheduler sigmas (0.1 to 10)." on Automatic1111's WebUI.

@P317cm 
Idk how to directly upload with exif data, i think imgur strips them so im just gonna paste them here.

Samples: https://imgur.com/a/23hVSxl

parameters

Young Blonde girl, Volumetric Lighting, Best Shadows, Shallow Depth of Field, Portrait Of Stunningly Beautiful Girl, Delicate Beautiful Attractive Face With Alluring Sharp Blue Eyes, Sharp Eyebrows, Enjoying The Scenery, Soft Smile, Closed Mouth, Lovely Medium Breasts, Slim Waist, Black Short Pleated Skirt, Thigh Gap, Black Thigh-High Stockings, Late Morning, Sunlight Shining Through Window, Floating Dust Particles, Hallway, Looking Out At Window, Standing, (Solo:1.3)
Negative prompt: (Worst Quality, Low Quality:1.4), Poorly Made Bad 3D, Lousy Bad Realistic
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 696983948, Size: 512x512, Model hash: 61e23e57ea, Model: mixProV4_v4

postprocessing

Postprocess upscale by: 4, Postprocess upscaler: 4x-UltraSharp

extras

Postprocess upscale by: 4, Postprocess upscaler: 4x-UltraSharp

You can see that compared to some images in the gallery the face area is somewhat poorly detailed, idk if thats just random or theres a reason for it, VAE im using is kl-f8-anime2.

I'd like to mention that ddetailer doesnt currently work in automatic111's webui, it just doesnt load it wheter you install from the extensions or manually.

@P317cm Alright it seems that ddetailer you linked is working, just to understand it correctly, do you put the prompt in the ddetailer section instead of the main prompt section? Or do you only put specific things in it? I never used it before

The inbuilt merger has limitations and you'll often end up with what you've been getting, or sometimes with some exceptions of models merging together very well. You'll want to look into block merging and read the information on this page about it. https://github.com/hako-mikan/sd-webui-supermerger, good luck!

@skordean Thank you for your tip skordean. I've creating a mix with the information there, you can check it here:
https://civitai.com/models/51711/icerealistic

@skordean o there is a video tutorial about merging loras to checkpoints using supermerger? i think im too dumb to follow a text tutorial

i also new. and has same problem...i try do this, finally slove it...but i don't know which one works

1. down new model v45colorbox

2. click Script and select "Detection Detailer"

3. click settings at top. select "Stable Diffusion" at left， and at right page . "SD VAE" section, select "kl-f8-anime2.ckpt"(i guess this work)

setp 2,3 need install somethings, it shows in home page

@KUMO233 hi, I have nothing in my list of Scripts after installing ddetailer, what should I do then?

Your problem is probably VAE. I think that VAE from creator isn't the best for this model. I personally used so-called "Blessed vae" (sorry, don't remember exact source of it) and it fixed an issue. In 99% cases of desaturation/overexposure in my generations VAE was the issue. The best will be download many different VAEs and compare the results, then bake-in your favourite into the model (Checkpoint Merger > No interpolation)

Lower your CFG scale. Excessive prompts with too many Loras, embeddings or hypernetworks can wash out your images. If you drop your CFG under 10 you should be good. If it's still fuzzy when CFG is 2 or lower then your prompt is too complex and needs a compiled solution like a better trained model.

+1

+1

+1

which vae was it?

@rozeolifant13 

You need a Vae like kl-f8-anime just type that in google