Trained on 41 images. You'll get decent results with any aspect ratio, though I mostly only used 512x512, 512x768, and 768x512.
Strengths
Produces consistently faithful recreations of the character when using the recommended settings below. This lora is also fairly flexible, allowing you to choose which aspects of the character you want to include or exclude in your image, as well as comfortably handling a wide variety of poses and situations.
Shortcomings
Struggles with hands and eyes in some situations, and doesn't seem to play nice with some of the more common checkpoints. Currently it also has trouble with Byakugan eyes, the shinobi outfit, and the white frills on the skirt, but I plan to make improvements to these areas in future versions.
For best results I recommend:
Use uzumaki_himawari in the beginning of the positive prompt.
To prompt accurate character details use: (lines on face), short black hair, blunt bangs, blue eyes, ((yellow hoodie, pink skirt)), black thigh highs, red boots,
Using this LoRA with CyberAlchemist's Animix - Anime Screenshot-like Style Mix LoRA
Use a lora weight around 0.5-0.7. Going over that seems to produce bad results.
Use Highres. fix with these settings
R-ESRGAN 4X+ Anime6B
Hires steps 0
Denoising strength 0.7
Upscale 1.5
Following DitamAI's No more color contamination guide to make sure the colors of the clothing stay consistent. I only isolated yellow, pink, and black (for the hoodie, pleated skirt, and thigh highs).
Use Clip skip 2
I plan to continue improving this LoRA with future versions.
If you have suggestions let me know.
Description
FAQ
Comments (7)
Why have a lora if you are going to tag every aspect of the character anyway?
??? so you can swap things out ???
it's better than character loras where all they do is tag it as the character so you have to jump through extra hoops because you they have accessories you don't want on them but it's so hardcoded to have those accessories that you have to lower the lora strength but then they stop resembling the face/eye/mouth shape. Like imagine trying to set up a widowmaker image but every person who makes a widowmaker decided that her suit didn't needed to be tagged so every image you try to make with her always includes it instead of letting you switch the outfit to literally anything else
@whyusecivit you can swap things out even if you don't tag everything, you just need a decently diverse dataset. And this is way better because when tags are overrepresented in the dataset your concept will bleed out into those tags, making results worse.
Any chance you can share the Ametrine Blend pleaes? Seems to be gone from everywhere.
While often not compatible with most 2D checkpoints (just like most naruto based models somehow), this model tango with some 2.5D checkpoints like no other.
Honorable mentions would be the Darksun (preferably v5 or v4.1) and JAM v1.8
Please share your ametrine blend model, i cant find it in any other site
Ametrine isn't my model. The creator did however upload a new version of the model here that I recommend checking out: https://civitai.com/models/57188/ametrine-redux
If that version doesn't work, then a legacy version was uploaded by a different user here: https://civitai.com/models/137746
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.