v3.0: This version was created by training version 1.5 on over 500 images with 92000 epochs. For this purpose, efforts were made to achieve SDXL quality. It gives better results at higher resolutions. Does not require VAE. Clip skip 1 and 2 give good results. I recommend using after detailer to enhance faces at low resolutions.
This model was created by mixing several different models with each other after patient and hard trials. It is the perfect choice for creating realistic looking anime pictures. You can also create more linear images by entering the correct prompts.
As a sampling method, Euler A provides excellent results.
CFG Scale must be at least 7 and Sampling steps must be at least 20.
I recommend using "4x-AnimeSharp" or "R-ESRGAN 4X+ Anime6B" as upscaler.
I usually use Clip Skip as 2 but 1 also gives good results
"Bad-hands-5", "bad_pictures" and "EasyNegative" are great for negative embedding.
I generally prefer to use hires fix instead of using restore faces. Restored faces often damage the expressions on the faces and the originality of the eyes.
I recommend adding negative embeddings and using hires fix to get decent results. Let your Hires fix steps be at least half of your sampling steps. Believe me, it will be worth it.
I can't do model training yet. For now, I'm just merging. I download gigabytes of files from civitai every day and now I have 5TB of data :) on the other hand, I'm trying to learn. I am trying to get proper results by mixing the appropriate loras and checkpoints with each other in certain proportions and I share the results that I find successful with you. Sometimes I don't remember what materials were included in the final model because I mixed too much. I guess I'll have to keep a long list to get around the problem. Very soon I will start to share the loras that I have produced myself. Sorry for my english because i don't know english :)
My work will continue...
Description
Same version but pruned one...
FAQ
Comments (6)
I'm telling you: i was very intrigued by "BeautyFool" and I tested the v1.1 whith great expectations. Unfortunately, at the present, is not the checkpoint for me. The reason is the heavvy negative embedding it needs to work. Without it, it would produce very basic faces, improved to fantastic results thanks to negative embedding. "Where is the problem?" you may ask. Unfortunately, this negative embedding tends to modify the intent of the prompt to standard results. For example, if you change age or breast size, the negative TI will restore them to the standard 'teen looking' of your gallery. Fantastic looking but is not what I asked for. I'm not saying you checkpoint is not good, and I still think your demo gallery is amazing... but this way to do is not for me... If, in the future versions, you may find a solution about this, I will happy to know!
After your "My work will continue...", please write something about the versions...
For example what's new or improved in your opinion...
can you train this checkpoint more to let people use character loras? i have to set weights insanely wide down to get your style, but those characters look nothing like they should. only how your checkpoint is trained....
The new pruned version 1.2 galvanized me to make more experiments. Also, I've searched about "negative TI". Have a read about this comparation, i'm sure you will find interesting:
https://civitai.com/models/56519?commentId=158376&modal=commentThread
This teached me to avoid adding more neg-TI, and eliminate "easynegative" with more control over the result!
you can be sure i will test it according to these directions, thank you.
Great model! I am pleased that you are working to improve it even further. Keep up the great work!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.