⚡️ Intro
This model is mainly intended to create nice looking semi-asian girls (NSFW&SFW).
Of course, you can try generating other things besides women 😎 Feel free to share Your work 💪
Inspired by QGO - PromptingReal. I love quality of nude pictures from that model 💙
So I merged it with GirlMix and RealDosMix using sd-webui-supermerger to get semi-asian results and better skin colors.
I also tried to merge it with many (≈15) popular realistic models. But I didn't get satisfying results🙄
So... click ⬇️ Show More ⬇️ link below to read some useful stuff, have fun... and try not to spend all your money for renting GPU's 😄😄😄
⚙️ Recommended Settings
This is the best setting for realistic faces, skin, details of clothes, details in background, etc.
CFG Scale 5-8
Clip Skip 1
You can also experiment with such settings. However, the images will be of lower quality.
CFG Scale 4-10
Clip Skip 2-3
Sampler DPM2 a Karras or DPM++ 2M SDE Karras // alterntively DPM++ 2S a Karras or DPM++ 2M Karras
Sampling Steps 35-60 for tests, 45-100 for the final work
VAE vae-ft-mse-840000-ema-pruned
You just need to do experiments and find your sweet spot. Check my XY plots for inspiration.
✍️ Recommended workflow
Set picture size to 540x960 (for portraits) or 960x540 (for landscapes). Don't go above 1024px!
Put nationality into positive prompt, for example
(Irish girl:1.3)or(Malaysian woman:1.3)to modify face shape and skin tone. You can read more about this in my article.Use negative embeddings or good negative prompt. You can find popular embeddings here. In my latest works I found that You don't need any negative Embeddings, all you need is to have good negative Prompt. Last one I stole from somebody and I don't remember from who (but shhhh😋)
Use Lora's for better image quality (read below)
Use 8x_NMKD-Superscale_150000_G or 4x-Ultrasharp for Upscale (strenght 0.2-0.4)
(optional) Put age into positive prompt. Be specific. Use keywords like:
XX years old,girl,woman,old lady,elderly woman.
Combining these words will give you the best results. Examples:photography of German girl, 20 years oldphotography of Italian woman, 40 years oldphotography of Spanish woman, very old lady, elderly woman
More info here.Do not use Celebrity / face morphing Textual Inversions because in most cases it will make your skin unnatural, pale. Better find good Lora and use it instead.
Do not use
Style of artist Xor...by artist Name Surnameprompts if you expect realistic skin.
✨ Recommended Lora's
Photoblend (play carefully with this - shadows and brightness may kill your work)
LowRA (use it for professional studio photos, for non-studio use about 0.2-0.3)
Shiny oiled skin (use Only for nudes, otherwise it can make your faces ugly)
🧾 Recommended Embeddings (Textual Inversions)
NSFW Embedding (awesome TI to keep girls naked or clothed)
🌏 Nationalities (ethnicity)
I started playing with nationalities after finding this Reddit post. Also I wrote an article about it.
Some nationalities produce sexy face outputs. You need to find your favourites.
Currently My favourites are: Caucasian, Malaysian, Iraqi and Grenadian .
For tests you can use sd-dynamic-prompts and this prompt.
For best results use Clip Skip 1 and keep in mind that prompting nationalities may affect clothes and backgrounds (if you don't specify them by yourself).
👎 Known weak points of this model
Almost impossible to generate tongue 👅
Almost impossible to generate realistic skin and body of nude grandmas 😦
Fingers 🖖 better hide them behind (common issue in many models)
Imperfect nipples from the side 😭 (common issue in many models)
🌫Outro
I hope this model will be beneficial to someone among you. If you enjoyed playing with this model, please leave a review and like other users' photos. Best regards 🖐
// all models and images were creted on Vast.ai servers (using RTX4090)
// all my images were generted without inpaiting and without ControlNet
// pruned model versions created with stable-diffusion-webui-model-toolkit
// if you have problems with my English - contact with my AI secretary 🧻
Description
This version will produce more asian-like women.
FAQ
Comments (9)
Great model, very flexible. Can generate people both young and old realistically.
Congratulations, a really great model!! I also tested using 480x480 (so it's a lot faster to create a heavy library by using wildcards), with minimal bodypart errors. Then I select the best ones and use IMG2IMG with ULTIMATE SD script (1280x1280 and beyond), with DENOISE 0.40 or less, and the same prompt. But, If the 480px sample needs correction, I skip ULTIMATE SD, then I fix it, be with a solid color and/or SMUDGE tool, then I save it and use IMG2IMG with Loopback scaler script (960x960), DENOISE 0.35 or less, 1 loop only instead of 4 loops, and if I like it, then I go to ULTIMATE SD.
Thanks for review and tips. Most of credits goes to authors of models I used to create this one 💙
Hi there,
Could you consider uploading this model to "tensor.art" as well?
It's a great completely free-to-use website for generating AI art using Stable Diffusion,
However, since it is a website hosting the tool that means you can only use whatever resources getting uploaded on that website.
Website have Virtual Credits. This model can not be used for Commercial purposes. For now I don't have time to take a closer look on tensor.art
@AddictiveFuture I understand, however as someone who uses it a lot I can tell you that those ''virtual credits'' are 100% free with literally zero way to buy them or get them in exchange for money, you get 100 credits daily without having to do anything.
The only way you can actively get more credits is by sharing the website, but not an any monetary way :)
Excellent work here! Your model is awesome.
Thanks 👍
First and foremost, this is not a checkpoint so much as it is a work of art. It's incredibly versatile, and the only limitation seems to be the imagination of the user.
Thank you for sharing this.
Greetings from the land of the rising sun.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.








