LoCon of Mejiro McQueen (umamusume).
Please install extension on WebUI.
LoCon, LoRA for Convolutional Network , is also kind of LoRA.
Sorry for late 😅. I was late to experiment with LoCon.
All of my works is free. If you'd like to support me, feel free to buy me a coffee. 👍
Recommend options
LoCon weight around 0.6
Trigger words
base
mejiro mcqueen \(umamusume\)
Racing competition clothes :
green bowtie, armband, long sleeves, frilled sleeves, frills, striped shirt, black coat, pleated skirt, black skirt, kneehighs, socks, black footwear, lace-up boots, cross-laced footwear
End of Sky :
official alternate costume, mini top hat, mini hat, white headwear, braid, ascot, clothing cutout, shoulder cutout, short sleeves, white gloves, cropped jacket, white jacket, belt, navel, midriff, short shorts, white shorts, white footwear, knee boots, lace-up boots, cross-laced footwear
(optional) feathered wings, feathers
Summer clothes :
official alternate costume, low twintails, hair ornament, hair flower, necklace, bracelet, green ribbon, swimsuit, white dress, bare shoulders, off-shoulder dress, frills, frilled dress, green bikini, sandals
(optional) see-through dress
Tracen school uniform :
white thighhighs, tracen school uniform, purple shirt, pleated skirt, puffy short sleeves, white skirt, puffy sleeves, summer uniform, frilled skirt, sailor collar, sailor shirt, miniskirt, frills
Casual clothes :
neck ribbon, pinafore dress, sleeveless dress, blue dress, white shirt, long sleeves
Use animefull based model
Settings
DPM++ Series(SDE Karras, 2M Karras, etc.)
about 20 steps, CFG scale 3.5~6.5
CLIP skip = 1
Use highres. fix to get higher quality image
Upscaler : Latent series
Denoising strength 0.50~0.65
LoRA training info
trained on sd-scripts by kohya_ss. Thank you a lot!
based on Animefull-pruned
247 images of mejiro mcqueen
4x low quality 102 images
8x medium quality 72 images
16x high quality 73 images
=> 1 epoch = 2152 images
replaced character feature tags with mejiro mcqueen \(umamusume\)
horse girl, horse tail, purple hair, purple eyes, etc.
resolution : 768x768
rank=alpha=16
conv_rank=conv_alpha=16
batch size = 4
other training settings is included in training_data.zip
This LoRA may be compatible with animefull-based models (ex. AbyssOrangeMix Series...)
All uploaded images are generated by AOM3A1
About LoCon
LoCon extends LoRA to almost all block of SD models.
LoRA only trains on Transformer block.
LoCon trains on Transformer block and Res block.
So, LoCon contains more parameters than LoRA.
IMO, LoCon is powerful when using highres.fix especially in Latent upscaling.
I tried to several different conv rank for LoCon, It comes out that the higher rank, the higher quality on hires.fix.
I think it's worth it to grow in rank size(file size).
And some people experiment with mathematically resizing LoRA.
If you interest in this, visit here.
Thank you KohakuBlueleaf for introducing LoCon.
And, special thanks to my best friend. DH.