Sabrina - Anima v1
My first ever LoRA!
This is my first attempt at training a character LoRA for Anima, using Sabrina as the test subject.
I trained this one through Civitai while I was figuring out the process. It's decent, but I've already spotted a bunch of things I'd do differently next time, particularly with dataset size, pose variety and outfit variation.
I'm planning to make a whole bunch of LoRAs for characters from older cartoons while I learn and improve my training process, so I'll probably come back and make an updated Sabrina LoRA once I've got more experience.
For now, enjoy my first attempt!
Trigger Word
sabr1nan1m
Recommended Prompt
highres, masterpiece, best quality, ultra-detailed, toon_\(style\), sabr1nan1m, 1girl, solo
Then add whatever outfit, pose, expression, camera angle, etc. you want.
Training
Base model: Anima
Training platform: Civitai
LoRA Dim: 32
LoRA Alpha: 32
Training steps: 1500
Dataset: 32 images
Best results: around epochs 3-4
This was a learning run, so these aren't necessarily the settings I'd recommend as the "optimal" configuration. I'm documenting them so people (I) know exactly what was used.
Notes
This LoRA was trained primarily to reproduce Sabrina's character appearance while allowing her clothing, poses and expressions to be changed.
The training dataset feels relatively small, so some poses, outfits and details may not be as consistent as I'd like. An updated version with a larger and more varied dataset is planned.
Original Outfit
1girl, solo, sabr1nan1m, thin waist, blue eyes, pink and white striped tank top, single pink stripe, pink skirt, hair clip, pink hair clip, orange star earring on her left ear, green moon earring on her right ear
Sabrina's orange star earring is on her right ear and her green moon earring is on her left ear. It likes to swap them or double them up
Recommended Settings
Strength: 0.7-1.0 as far as ive seen
CFG: 5 or 6
Sampler: Euler
Scheduler: Simple
Description
Comments (1)
Good to see her done since shes been on the distant "someday" list so I'll dump a bunch of advice from one of the training grognards because its less work than actually making the model myself (and because I wanted this kind of advice when I started).
For tags: You don't need to tag body attributes if they're consistent on the character between images and you aren't trying to remove them in gen by excluding them from the prompt (the same way you don't need to explain the shape of jar or every curve of hair: Its learned as part of the LoRA by having the patterns be consistent). Thus "thin waist" is pointless at best and likely a distraction for the program at worst. Likewise you don't need/want quality tags in your training tags. You do however want "midriff" or "navel" as applicable (Confusingly those two are actually mutually exclusive: Navel is when the umbilicus is visible, and midriff is for when no umbilicus is visible but the area between pelvis and waist is) in your tags since they're both expressing the visibility of bodypart and clothing shape.
I recommend matching existing Danbooru tags as much as possible without inserting extra baggage like colors: You don't need to specify color unless you've got a mix of color and black+white images in your dataset, or the character frequently wears different pallet swapped outfits. "single stripe" and "mismatched earrings" are both recognized Danbooru tags and work well, the second in particular would be needed for any Illustrious port.
As for the dataset size: I've long recommended 20 as the bare minimum to even bother as more than an experiment, and past ~30 (for a single outfit character) is the point where you should start caring about average data quality more than raw quantity (that is, if you've got 30+ good images you should start removing low quality images from dataset if they're not needed for some outfit etc.). The whole series is officially free on YouTube, and its easy enough to get screenshots there (use the < and > keys to framewalk, double right click to save frame without UI) and thus it should be easy to get a large enough dataset (though beware the old uploads were stretched to 16:9 and would need some processing to return to normal). Given she dresses up every few episodes, I'd recommend adding some extra outfits on top of this if you can (e.g., the "trendy" clothes from Shrink to Fit) as long as you can avoid tag overlap. Generally even just a couple official pics in an alt outfit (even if not enough to learn it, but more is even better) will greatly enhance how well a model understands a character's body because it gives the trainer a way to more clearly distinguish the pattern of the character's body from the pattern of the character's normal clothes.