2023/7/30 experimental update
Upload sdxl version trained using sdxl v1.0 base model. Running SDXL on webui requires at least 8GB of CUDA memory. Currently, the generating quality of this version is lower than 6.0 anime version.
Update dataset. Remove trigger tag, so there is no need to input trigger tag when generating images using the latest version. Retrain the model using new craft.
Clean up the old versions and retain the versions with good generated results.
Update dataset. Adjust trigger tag to 'fashi-girl'.
BTW, make sure set this option in 'Stable Diffusion' settings to 'CPU' to successfully regenerate the preview images with the same seed.
Update dataset. Slightly optimize body shape. Try to make the face more alluring. v5.4 version is conventional LoRA model.
Upload Lycoris version (v5.3). Remove hyper-realistic images in dataset that may cause bad results. Try to make face more natural and beautiful.
Due to the instability of Lycoris caused by sensitivity to learning rate settings, the updated model (5.2 version) is still conventional LoRA model. Use images published by some users as dataset to train the model. Fix the compatibility problem of non-NAI-based checkpoints. Now arbitrary anime model with NAI's VAE or kl-f8-anime2 VAE can also generate good results using this LoRA, theoretically. This version is a bit overfitted that will be fixed next time.
2023/3/24 Experimental Update
New model is trained using Lycoris method (all versions before 5.1 version are conventional LoRA model), pls install this extension to use this new model in webui. Adjust trigger tag to 'fashi-g'. If some problem occur during generating pls use old version.
Update dataset and model. Decide to upload 5-epoch version whose color saturation is a bit better than higher-epoch version after careful consideration. Adjust rank to 128 to keep perfect face. The trigger tag is unchanged. The preview images are generated using latest version of YesMix and EasyNegative embedding.
Reduce loss to ~0.11 using Lion optimizer. Btw, if you use Hires. fix, 0.55~0.6 weights of Denoising should be enough, otherwise the face may be slightly deformed. Btw, the preview images are generated using latest version of YesMix and EasyNegative embedding.
Update dataset and model. Add trigger tag 'fashi-girl' to improve generating probability. Adjust rank to 32 to balance file size and quality.
Prune tags to make the description more accurate.
Merge previous LORA model that is easy to generate nice images to slightly improve color saturation and optimize face. Adjust rank to 32 to balance file size and quality. The preview images of new version are generated using latest version of YesMix.
Adjust dataset to fix bad philtrum. Reduce training epochs to fix deformed limbs. The preview images of new version are generated using latest version of YesMix.
Update and balance dataset to remain the images of mature female that fit personal aesthetics. New version of LORA model is trained using Lion optimizer discovered by Google Brain that is purportedly better than Adam(w). The preview images of new version are generated using latest version of YesMix.
Update dataset and model. The preview images of new version are generated using latest version of YesMix.
Clean the dataset to remove the images of unattractive characters. Use offset noise to improve ability to generate lighter/darker images. The preview images are generated using latest version of YesMix.
Dataset is updated, so model is updated. Reduce the model file size to 1MB. The generating quality is unchanged.
This is an experimental LORA model trained with about 100 images of fashionable girls that fit personal aesthetics. Therefore, this model is named as "Fashion Girl". Due to the small-scale dataset that are composed of realistic/photorealistic images, some output images will remain anime style. The 784mb VAEs (NAI, Orangemix, Anything, Counterfeit) are recommended. 0.6~0.8 weights should be enough. This model will be continuously updated as the dataset is updated.