V3, tryed a new training method





workflow: https://civitai.com/models/1088678
About my celebrities // character loras
The dataset used to build this lora is a mix of celebritie face and character in context images.
To get the celebritie only, you can use the trigger word, and not the other one: ohwx woman
To get the character, you can use all, or a mix of ohwx woman along with other trigger words mentionned in the description.
Notes:
***To get a better likeness, use higher weight. I see 2 images in the dataset, that Supergirl is not 100% full likeness that may have caused that.
***For some reason, if you don't prompt the background or context, you may get a red carpet image where there is a marking background. Just prompt your background.
To get the best result
Guidance: 2.2-3 (I start with 2.2 as it give me best result, but if the image is deformed or color are washed, I rise up that number until full satisfaction)
Steps (dev): 30-40
daemon detailer (lying sigma sampler): factor: -0.02, start 0.06, end 0.75
Resolution: Upscale the latent by 1.25 or 1.5 you'll get awsome result. (take longer time but worth it)(I usualy use 1.4, that give me outstanding result but sometime it add a strange color frame around the picture)
Enjoy!
V3, tryed a new training method
1. Quyền đối với các mô hình được đăng lại thuộc về người sáng tạo ban đầu.
2. Người sáng tạo gốc muốn xác nhận mô hình vui lòng liên hệ nhân viên SeaArt AI qua kênh chính thức. Nhấp để xác nhận
