All the versions are different and I cannot tell which one is better. You can try or view the sample pictures first. In most cases hires fix is needed to get the quality of the sample image.

updated 3.0a. This is the version that only use anime pictures and hard-trained. It is a "hard" model that trains a lot. Which means it is very hard to control and aimed at restoring character. So 3.0b is still the version I recommend unless you are not satisfied with the images that generated by that version or you can use LoRA Block Weight. There are some example pictures generated by using LoRA Block Weight and you can see the performance.

Use a new way of training that do not need AI generated pictures anymore. Update 3.0b,which is an specialized version that use lower epoch which trains slightly. The base version 3.0 has some problems but will release soon.

Based on S2 and S3 anime's character design to make the character stable.

Works well in the case closer than fullbody, because there is no good fullbody materials in anime. Try to use a low weight and/or hires fix in that case. It still works well in a very low weight if the prompts are used.

Works better in the pictures width<height. For that is the main part of the materials.

For the reason above and for nsfw use, some nsfw pictures trained by the original version is added. But it causes some new problem such as the slightly mismatched color of eyes. And I find that nsfw problem in the original version can still be handled by using lower weight. So I may upload the original non-nsfw version(almost only based on anime) as well.