Westmix V.1, Photo Realistic Model for Beautiful Western Girl

Hires : ESRGAN x2/30/0.3, clip skip: 2, Extensions: dynamic threshold, ddtailer (Just an example.)

Several days ago, I was bored, so just for fun, I quickly made a model and shared it here. But surprisingly, many people showed interest, so this time, I stayed up for three nights, did thorough testing, and now I'm uploading the official version, v1.

1. Update Summary

The first official version, v1, is designed to generate beautiful Western female images quickly, depending on people's preferences and computer specs. It can generate high quality images with shorter prompts, a low sampling step, and a fast sampler (without LORAs, embeddings...etc). I also set high thresholds for prompt and setting values, so skilled folks with high-performance computers can create top-quality images by adjusting prompts and values. At the same time, I intentionally biased the default face of the model towards Western person while not allowing the model to generate exaggerated western woman's face (for Asian users' tastes). Lastly, I made sure to maximize prompt recognition, so it has a very high success rate compared to other photo-realistic models.

2. Change in Westmix V1.0

1) The biased Asian face as the default face has been eliminated, and the need to add "Asian" in negative prompts is removed.

2) To generate high-quality images that closely resemble real photos, you don't need lengthy and messy positive and negative prompts anymore. (ex. 4~7 neg. prompts)

3) Significant improvements have been made to generate top-quality results even with fewer sampling steps and lower configuration settings, catering to users with less powerful computer specifications. (Also With DPM ++ 2m Kerras, which is much faster than SDE 2m Kerras, you can now generate top-quality images)

4) The face generation process has been enhanced to create faces that closely resemble real photos and can produce more various faces than the previous testing version V0.

5) The prompt recognition has significantly improved as the number of prompts required to achieve high-quality results close to reality has been greatly reduced.

3. About Sample Images

Like in the previous model, all sample images are without any corrections or selections. I uploaded whatever Mr. AI produced. This means that you will see odd fingers, hands, toes and feet on some of them.

A few top-quality images are without prompts and settings again. I want you to play with many models yourself and be good at adjusting setting numbers and prompts for photo realism because it's really fun!

Sample images without exifs : only one offnoise related lora used. No embeddings, no loras. (If they look good to your eyes, it's the power of prompts and numbers. FIND THEM by playing around with it!!)

I put a bit of explanation on each set of sample images, so if you are interested, read the post titles and descriptions.

4. About Extensions

1) Use at least DDtailer and Dynamic Threshold extension. They will help you generate top-quality images. (DDtailer is a must to use when you generate a far distance long shot.)

5. Sampler

With DPM++ 2M Kerras, which is much faster than SDE, you can generate top-quality realistic images. You can find some sample images I uploaded. Find the best setting and prompts for the model and enjoy the best quality images.

With DPM++ SDE, you will see faces are slightly different from that with 2M Kerras. They are more 'western-beauty-standard-like' than in 2M Kerras. And, of course much slower too.

6. NSFW

Unfortunately, I am not interested in NSFW at all and I don't even know how my model does for NSFW. I am rather interested in figuring out how this beautiful tool can be applied to everyday life.

7. Things to improve in the future

1) Personally, I really care about the facial part. It's still not as photo-realistic as I'd like it to be. In the upcoming updates, I plan to maintain v1's power while enhancing the realism of the faces.

2) A tiny bit unstable when everything exceeds its threshold. I don't think you will find the situation where it does because it's just from how I test a model. I will have to improve it anyway for a perfection.