This Lora produces a mix of 3d and realism. I have trained this lora over 80 different times, getting it closer and closer to perfection. The dataset that I used to make this model took me a very long time to make as well, perfecting every single photo in photoshop, one by one.
It was not trained on the standard resolution of 1024x1024 ether. I trained it in AI Toolkit using the resolution of 2048x2048, (the same resolution as each image in the dataset). I honestly don't remember the learning rate I used for it. I know I used Adamw and turned quantize off, it took a big boy to train it on runpod.
The images come out with very smooth skin. You will also rarely get the rabbit teeth when using high strengths or the butt chin. If you want the images to have skin detail as well, then you just run it through a second Sampler, you can inject latent noise into the second sampler as well to get a better yet inconsistent effect. I don't use this lora in the detailer nodes, just the first generation as the skin is already super smooth.
I wanted the display images shown above to have used relatively simple prompts. No other lora was used in the intitial generation for any photo, however I did run some through detailers and such. At the end of the display images you will see raw batches straight from the Ksampler.
Play with the strength. Personally I like it above .5 and normally use it with .8. With low strengths, it just looks weird.
These are the prompts I used for the display images, it will be pretty easy to determine which image came from which prompt.
Prompts for display images (without the numbers obviously):
#1. Create an intimate, photo in 8K resolution capturing a redheaded young woman's face. Heavy snow falls dramatically around her, with individual snowflakes visible and glistening as they catch the light. Powerful moonlight streams through the tree canopy.
#2. A 64k portrait of a semi-realistic looking young woman. She has flawless and shiny skin that glistens under the light. The background is spectacular with intricate details throughout. The overall photo is has Cinematic lighting with a polarizing filter, and everything in sharp focus.
#3. A young woman, standing out in the rain, smiling. She has bright red hair, tied up in a ponytail. She has soft eyes. The photo has cinematic lighting.
#4. Create a photo of a young woman with a cinematic background and strong natural lighting.
***I know the filesize is big, I also do not care.
I have over 80 versions of this lora, not epochs, versions. However, only a couple were trained in this manner and each one trained on runpod has this file size. Most were trained on Civit.
I do have a few versions in smaller sizes, I do not believe the quality is the same but if this lora is popular I can upload the smaller sized versions that were trained on Civit.
This Lora produces a mix of 3d and realism. I have trained this lora over 80 different times, getting it closer and closer to perfection. The dataset that I used to make this model took me a very long time to make as well, perfecting every single photo in photoshop, one by one.
It was not trained on the standard resolution of 1024x1024 ether. I trained it in AI Toolkit using the resolution of 2048x2048, (the same resolution as each image in the dataset). I honestly don't remember the learning rate I used for it. I know I used Adamw and turned quantize off, it took a big boy to train it on runpod.
The images come out with very smooth skin. You will also rarely get the rabbit teeth when using high strengths or the butt chin. If you want the images to have skin detail as well, then you just run it through a second Sampler, you can inject latent noise into the second sampler as well to get a better yet inconsistent effect. I don't use this lora in the detailer nodes, just the first generation as the skin is already super smooth.
I wanted the display images shown above to have used relatively simple prompts. No other lora was used in the intitial generation for any photo, however I did run some through detailers and such. At the end of the display images you will see raw batches straight from the Ksampler.
Play with the strength. Personally I like it above .5 and normally use it with .8. With low strengths, it just looks weird.
These are the prompts I used for the display images, it will be pretty easy to determine which image came from which prompt.
Prompts for display images (without the numbers obviously):
#1. Create an intimate, photo in 8K resolution capturing a redheaded young woman's face. Heavy snow falls dramatically around her, with individual snowflakes visible and glistening as they catch the light. Powerful moonlight streams through the tree canopy.
#2. A 64k portrait of a semi-realistic looking young woman. She has flawless and shiny skin that glistens under the light. The background is spectacular with intricate details throughout. The overall photo is has Cinematic lighting with a polarizing filter, and everything in sharp focus.
#3. A young woman, standing out in the rain, smiling. She has bright red hair, tied up in a ponytail. She has soft eyes. The photo has cinematic lighting.
#4. Create a photo of a young woman with a cinematic background and strong natural lighting.
***I know the filesize is big, I also do not care.
I have over 80 versions of this lora, not epochs, versions. However, only a couple were trained in this manner and each one trained on runpod has this file size. Most were trained on Civit.
I do have a few versions in smaller sizes, I do not believe the quality is the same but if this lora is popular I can upload the smaller sized versions that were trained on Civit.