
V1.0_IL
IL models can generate her without LoRA but will either have the modern art style or some mismatches in her dress. So this version is made specifically to address my Pet Peeves.
By far IL LoRAs are the most accurate to the original artworks in both details and style. (Surprisingly capable of other styles as well)
This version did a bit overtrain on her royal dress. Causing her other attires to require more specific prompts.
V1.5
I forgot to set the caption extension in Kohya. So the previous version is actually trained without any caption.
This version should be less overtrained and easier to work with other styles/LoRA.
While it is now required captions to be more specific. But also there are way less background artifacts so negative is no longer required.
V1.0
A character from Brighter Than the Dawning Blue (夜明け前より瑠璃色な) series
For non-Pony LoRA. I recommend this person's LoRA https://civitai.com/models/268071/sdxl-character-9character-set
Character Tag
Feena Fam Earthlight, green eyes
(May get blurry eyes without the tag)
Costume tags
(Can add color/accessories to enforce more specific parts)
Dress
Princess, Dress, Tiara, Brooch, Elbow Gloves, Cleavage
Sundress
Sundress, Straw Hat
School Uniform
School uniform, Pleated skirt, White thighhighs, Bow, Emblem
Bikini
White bikini, Straw hat, Sarong
Negative Tags (Optional)
Glow, Sun
(The majority of training data have kind of glowy outline or a Moon in the background. Thus it tends to sneak it's way into the result if you try to have night sky and something similar)
描述:
训练词语: Feena Fam Earthlight
名称: Feena_SDXL_2-000095.safetensors
大小 (KB): 445789
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success