V5 has been released.
Added training images and revised dataset tags.
Changed training size to 1024,1024.
Eyes were always in a surprised state, so we have made it possible to suppress this.
Since the expressiveness has decreased, we recommend continuing to use previous versions if it is difficult to use.
This version also contains a lot of experimental meaning.
We recommend that CFG SCALE not be larger than 3. The pattern will be intense. 2.5 or less seems to be just right.
There are still cases where the hook breaks depending on the direction of the face.
If the hook breaks at a long distance, try using ADetailer.
If you have any opinions, please comment. If it can be fixed in the next update, we will fix it.
V4 has been released.
We have succeeded in reducing the file size.
We have reviewed the dataset and tags. Previously, the anime style was too strong, so we tried various things to reduce it.
I adjusted the dataset tag so that if you enter [anime color, anime style] in the negative prompt, the art style will fade away from anime style.
If you use only the trigger word [nosehook], the hook will be set in your nostril and will not stretch very much.
By using it in combination with [nostrilspullup,long nostrils], you can unleash your true potential.
I think the expressiveness of the prompts has improved considerably., but there are still some issues to be fixed.
If you have any suggestions for fixes, please comment. I will try what I can before updating the version.
It is normal for file release versions to differ.
V3 has been released.
I have fundamentally reviewed the training method and tried to reduce the file size.
After 10 tests, I have finally made it usable for release.
Our focus has been on reducing file size, so if you are concerned about a decrease in quality, please use an older version.
I divided the concept of nosehook and had it learn.
1.nosehook
2.nostrilspullup
3.long nostrils
By dividing these, the reaction to the base