2023/7/17 update
Upload the realistic version to generate images using arbitrary realistic models. Please make sure to set the clip skip as 2 when generating images. The generalization ability of the realistic version has been tested on multiple realistic models. See the preview images for details.
2023/7/16 update
Update dataset. Retrain the model using new craft.
2023/3/11 update
Update dataset and model. Reduce loss to 0.09 using Lion optimizer. If you want to stably generate good results, pls make sure your sampler settings are identical to the following settings (and make sure your clip_skip is set as 2):
eta for DDIM: 1 (0.67~1 should be ok)
eta for ancestral samplers: 1 (must be set as 1)
sigma noise: 1
eta seed noise delta: 31337
Introduction
A LORA model trained with ~28 images of Shiranui Mai by Akiryo. The trigger tags are 'akiryo-mai' and 'shiranui mai'. 0.6~0.8 weights are good. The preview images are generated using latest version of YesMix and EasyNegative embedding.
Description
Update dataset and model. Reduce loss to 0.09 using Lion optimizer.
FAQ
Comments (10)
NAI‘s VAE or Orangemix's VAE
@zakp Please tell me what model I want to use. I can use 3D AI like you. My AI is cartoon
@wop1359602 You can try to use Yesmix-v1.5 or Orangemix. Maybe can generate 3d effect.
Hello! I wonder why I use the same model, same prompt, same eta, same CGF, it has a beautiful face when I close up Mai's face, but her face always ugly when the photo is far away from her.
This is one of shortcomings of stable-diffusion. Try to use Hires Fix to generate higher-resolution images.
@zakp It's a little bit better than before, but also a little ugly, still cannot have a beautiful face like yours. 😅
The face is based on what ckpt models used, so it comes out ugly if we using different models, i hope you train mai face for lora as well