VERSION 2 IS HERE!

Stronger details, better shading, and more.

I trained V2 with double the steps and half the batch size, and used prodigy to handle the rates. It learned fairly good, but had to sacrifice some LoRAs as I used restarts and it burned a few sadly; that's why I have gaps in my comparison picture.

Make sure to choose the version you see fit. V1 leans towards flexibility on the expense of perfect ???????, while V2 leans towards the opposite.

Trained for PonyXL V6 on khoya-ss XL colab notebook.

Feel free to order your lora through this article: https://civitai.com/articles/6872/commisions-available-18-only

I noticed that Pony XL V6 is already trained on Doppel's art style, but it's not quite polished. So I made a dataset. Trigger word is (Doppel Style), and it is the same that is used to trigger the style nativly.

You still need to use the recommended VAE.

For better results, please use the prompting format I used in my showcase, which is tags. Using natural language won't be as flexible as using tags.

Note1: This LoRA is strictly made for PonyXL V6, it may not work with other models.

Note2: Use it with 0.8 to 1.0 weight and DPM+3 SDE KERRAS, it gives the best results. It's not available on Civit's free tier I guess, so make sure to download the lora and test it locally. Also you may need to run the result through a face detailer; All the showcase photos ran through one.

Note3: Using the Lora in a weight more than 1 (For example 1.3) will give you extremely accurate results if you aim for the original art style in a nearly 1:1 comparation, but it will rip off the flexibility of using spatial control nets, a canny net for instance.