AerialLand-Pony-PAseer

Base on Matrix-Realistic-Pony and RealPony, I trained Asian Faces in the checkpoint which cost ≈50 hours. It is hard to say this is a well trained checkpoint, because the model base mostly on two checkpoints I mentioned above. But its denoising logic has changed significantly.

When use 【realistic】 as prompts, the other models always give out bad results or deliberately weird facial shapes.

I tried to unblock and eliminate this problem, and from my limitated tests, this attempt seems to be successful in most cases, making the face more various generalizable.

So, now you don’t have to be afraid to use prompt words such as [4K] [HD] or [Realistic] [photorealistic]. And, notely, the default characters are more Asian 【if you want to make sure, please use Asian or Chinese】, but if you want other kinds of face, please use 【western】or country name (didn't have enough test).

Because it is essentially trained based on PonyV6 as the basic model, the picture will have a 2.5D texture. This is something that cannot be avoided with this checkpoint, and if you are looking for complete realism, this checkpoint may not meet your requirements.

But, anyway, please enjoy and do not hesitate to share your pictures.

If you like my work, you can support me a cup of coffee at https://ko-fi.com/asseer.

Thanks in advance.

Recommend downloading locally for image generation.

Please note: This model is based on the Pony XL engine and introduces a rich set of lewd images for the SDXL, soplease strictly abide by the website rules and regulations. Any images generated based on this model that violate community rules and laws are not related to the creator except for images created by the creator.

Prompt word rules: Positive prompt wordsmust start with【score_9, score_8】, and the rest is the same as a standard model. Negative prompt wordscan include【score_4, score_3】 to improve stability.

This model is based on Matrix-Realistic-Pony and RealPony, I added training for Asian faces, which took about 50 hours.

It's hard to say this is a fully trained checkpoint because the model is mainly based on the two models I mentioned. However, its denoising logic has undergone significant changes.

I found that when using 【Realistic】 as a prompt word, other models always give poor results or intentionally weird facial shapes (like a pancake face).

I attempted to address and resolve this issue, and from my limited testing, this approach seems to be successful in most cases, making the faces more versatile.

This means the model can accommodate more LORA and reduces the impact of native model logic on LORA.

Therefore, you no longer need to hesitate to use prompt words like [4K], [HD], [Realistic], or [photorealistic]. It's worth noting that the default characters are more Asian (unless you specify Asian or Chinese to ensure this), and if you want other types of faces, use [western] or a country name (not tested, efficacy unknown).

As it's essentially trained on PonyV6 as the base model, the images will have a 2.5D texture. This is inherent to this checkpoint, and if you seek complete realism, this checkpoint might not meet your standards.

Nevertheless, please enjoy and feel free to share your photos.

If you enjoy my work, you can buy me a coffee at https://ko-fi.com/asseer.

Thank you in advance.