Trained on a large dataset with images of Red Sonja and Conan.
Illustrious usage and notes:
Prompt for Conan: 1boy, conan, shirtless male, muscular, blue eyes, loincloth
Prompt for Red Sonja: 1girl, RedSonja, red hair, long hair, green eyes, armor, bikini armor, breasts, gloves
A lot of the tags above aren't necessary. You'll get the character with fewer tags added.
For Conan and Red Sonja in the same image, start with this prompt: 1girl, RedSonja, 1boy, Conan
If you add comic to the prompt then you'll get a style that's similar to what's in the comic books.
You can add comic cover for comic cover-style image.
For full prompt examples, look at metadata for the image uploads.
PonyXL usage and notes:
Prompt for Conan: 1boy, Conan
Prompt for Red Sonja: 1girl, redsonja, green eyes
You can make Red Sonja's design more consistent by adding redhead and bikini armor
If you add comic to the prompt then you'll get a style that's similar to what's in the comic books.
Female characters might tend to squint or keep their eyes closed, but if you specify an eye color then it doesn't happen.
Characters tend to frown by default. I guess no one is happy in the Conan universe.
For full prompt examples, look at metadata for the image uploads.
Description
FAQ
Comments (9)
Heh. Didn't see about the squinting part after the fact. Whoopsie! Just as well since I was going for scowling anyway. ;)
This Sonja looks gorgeous! from which comics or cover arts you pick for dataset?
The dataset is pretty huge: around 1887 images. The dataset is a mix of a lot of stuff: cover art, individual comic panels, fan art, and renders. The stuff from comics is from relatively new comics, as in the last 20 years.
@WizardMcDuck Damn that's a lot of pics :D Seems like you did the definitive Red Sonja lora lol. I'm also assuming you trained it local because in site trainer limits 1000 pic, great work, loved it.
@chrisss1 Yeah, I train all of my LoRAs locally. The version I uploaded took 6 and a half hours to be trained. My other LoRAs only needed about 20 minutes of training.
But, to be fair, the amount of images is complete overkill. I did it partially as a test to see if a massive dataset would lead to better quality, and while it's a pretty versatile model, I think the quality would be slightly higher if I curated the dataset and trimmed it to the best looking images.
@WizardMcDuck I agree. Also heard this many times, too much of pics just overkill, quality over quantity. I saw people making very good loras with only 25-30 image and they also trains locally. Pretty impressive ngl. It's all about settings i think. I want to give onetrainer a try but my gpu is weak, rtx 2060, not sure if i do fine with it.
Haven't really used this enough to know exactly what triggers it. If its this lora or model with animated focus. But it can put out close ups as well as comic like prints. The original quality of the comics werent super good on that model (animics) but with some refinement in other model (kegant) it came out pretty well. And well I like that, conan style comics that is.
If I recall, I did include quite a lot of comic panels in the dataset. Using the tag "comic" and avoiding other style keywords should be enough to make the model lean in that direction.
@WizardMcDuck yeah I noticed that, I have used it with other loras as well to make other chars get into the comics, its works pretty well. appreciate the lora alot!