Mallow - マオ Mao - School Uniform, Magical Girl (Pokemon Sun and Moon)
Trigger Words
mag3ca gir7l,
Trigger Words - School Uniform
mall7ow_un7form,
green eyes,
dark skin,dark-skinned female,
green hair,twintails, curly pigtails,hair clips,
yellow backpack,
gray school uniform,sailor collar, red scarf,
white socks, brown shoes,
Trigger Words - Magical Girl
mall7ow_mag3cal,
green eyes,
dark skin,dark-skinned female,
curly pink hair in high ponytail,pink flower hair hornament,
yellow earrings,
makeup,
pink cape, pink dress,pink wristbands,
pink socks,pink heels,
Description
FAQ
Comments (3)
It's all just awful and terrible. I really don't like the quality of the model, but oh well—someday I'll figure out what I'm doing wrong with the anima models, but in the meantime...
do you use lora dataset studio to curate your datasets? My workflow is currently lora dataset studio for web scraping > comfyui for batch editing the dataset with things like face autocrop (I prefer doing this to letting krea2 or whatever generate the faces based off of a sample image inside of lora dataset studio), denoise workflow, background generation for images with simple backgrounds, and upscale workflow for 2x and 4x upscaling (depending on original image size) > lora dataset studio again for further isolation of the character with things like cropping out watermarks, watermark removal, and finally, letting it chose a balanced set or a dynamic set using siglip2. Exporting that bank to an empty character dataset then choosing the "prose" captioning option and having it write the captions to text files > then I use that curated dataset in the "anima lora" program by sorryhyun and toggle on both image masking options to further remove letters and chat bubbles. I also use it to write danbooru style tags and have it merge those tags with the curated datatset's prose caption text files (afaik anima prefers the hybrid prose+tags captioning approach when it comes to its training dataset [the anima lora program will use the merged versions upon training if that setting is toggled on, by default it's off]). Personally, I check the "post_image_dataset" folder after running all of the caching and masking steps under the "Preprocessing" tab, before starting the lora training to make sure the dataset merge and masking went okay. I also prefer the tlora method of training (in my experience the results are usually more accurate). If you're already doing this, then that's great. It's just what I have found to produce results that look clean enough to me. I haven't yet uploaded a lot of my completed lora's here. But I've noticed anima is very sensitive to one or two images that don't align with the rest of the dataset's visual pattern. For instance, one of my characters inherited the wrong sclera eye color just from two images having that difference, even the though the dataset consisted of 60 images of the same character. If you want, I could share the comfyui workflow files (a few are my custom alterations). But I'm usually very busy on training llm and sd models, so idk how fast that would be.
@cardistartist I use the Civitai training program, im use forge,I used the basic labeling I used in Illustrous,Definitely the worst option

