Hey there! ChatGPT and I are back with a brand-new LoRA! This time, I’ve included Illustrious-XL, and I must say, I’m very satisfied with the results.
This LoRA is inspired by the Victorian era, making it perfect for creating Vampiresses, Maids, and Historical Costumes. That said, it also works beautifully with modern attire, adding intricate details that enhance the aesthetic of any image.
Dataset:
You can check out the dataset used to train this LoRA here: Crimson Maiden Style Dataset.
Realism & Strength:
This LoRA performs well in realism when used at medium strength (0.5–0.7). Higher strengths push it toward a more illustrated style, allowing for flexibility depending on your creative goals.
Recommended Configurations:
PonySteps: 40
CFG Scale: 7
Sampler: Euler Ancestral
Strength:
0.5 to 0.8: Integrates harmoniously with the model.
0.8 to 1.2: The LoRA takes center stage for a more pronounced effect.
Resolution: Works well with most resolutions, but 1024x1024 and 928x1232 are highly recommended.
Steps: 28
CFG Scale: 5
Sampler: Euler Ancestral
Strength:
0.5 to 0.9: Balances perfectly with the model.
0.9 to 1.4: Gives the LoRA more dominance.
Resolution: Compatible with most resolutions, but 1024x1024 and 928x1232 yield excellent results.
I hope you enjoy this LoRA! It’s a versatile tool for achieving stunning results, whether you’re aiming for realism or illustration. Let me see your creations—I’d love to check them out! 🎨✨
Description
Pony Version
FAQ
Comments (7)
Awesome style!
Doesn't seem to have an effect on my images (and the file is tiny, only 11MB?)
Can you share more info? What is your base model? Are you using ComfyUI, A111, Forge? By the way i appreciate your feedback as those are my first steps training.
Using A1111, tried a few base models, including NTR-Mix and NovaOrange
@bscientist I am using the Diffusers text-to-image method to train this LoRA, I realized that I need to adjust the rank parameter. Currently, I’m using the default value of 4. The rank parameter controls the dimensions that the LoRA will train on: higher values allow to capture more details and become more noticeable, also increase the size of the file. However, I want to ensure that the LoRA doesn’t override the characteristics of the base model. I’ll be working on a new version with this in mind. Thank you again for the feedback!
you may want to retag all of your images, and remove all of the underscores from your tags.
Right now i am using DeepDanboruu to generate the tags with a tag threshold of .5 i notice that maybe i can increase it and didn't realize about the underscores thing, will make adjustments for further versions, thanks for the feedback!!