New dataset mostly based on anime screencaps, focused toward the Tenchi version of Sasami. I intend to train Pretty Sammy version separately.








Used kohya_ss gui. Pony version trained on Civitai On-site Trainer.
Pony version:
Only trained on Tenchi version of Sasami for now, I figured to do a separate model for the Pretty Sammy version.
• Trigger Word: sasami
• Weight: 0.75 - 1
Keywords:
sasami, blue hair, long pigtails or twintails
Hair is not always consistent; if the hair color or style isn't coming through, add the blue hair/long pigtails/twintails keywords. Will investigate this more in the future. Also wider images will give her pigtails more room.
The classic anime style of the screencaps seems to be very strong, I would recommend using a lower weight along with adding "anime screencap" to negatives, which should help when adjusting style or using style loras.
Showcase images generated using ComfyUI, workflows although very simple should be available on each image. Tested on the following checkpoints:
autismmixSDXL_autismmixConfetti, cashmoneyAnimev10, tPonynai3_v51WeightOptimized
• Trigger Word: sasami or pretty sammy
• Weight: 1
Keywords
aqua hair, twintails, forehead mark
[Sasami]: hair bobbles
[Pretty Sammy]: hair flower, magical girl
Clothing helpers: chinese clothes, dress, skirt, long sleeves
To produce similar results I recommend EasyNegative, this negative embedding was used in the sample images. I'm also using Clip Skip 2 and hires fix.
New dataset mostly based on anime screencaps, focused toward the Tenchi version of Sasami. I intend to train Pretty Sammy version separately.
1.転載モデルは学習・共有目的のみで使用し、著作権は原作者に帰属します
2.モデルの認証は公式チャンネルでご連絡ください。クリエイターの権利保護に努めています クリックして認証
