Is there any approximate date for the full release? Also, a few questions about prompting.
I know the training time can vary quite a bit and unexpected things can happen but wanted to know if there's an estimated date for the full release, or to know if it's at least this year.
About prompting, it seems some booru tags, like the concept ones, can be pretty strong, in the sense that if most of the art for a concept has only been done by a few artists, the output style will lean to those artist styles and could include some things that happen to appear in those concepts often, like for example, columns of text in Japanese around the characters, despite having the respective negatives. I could name one or two nsfw tag examples, but not sure if that would be allowed here. A normal example would be the way that if an artist often has a watermark or their name on the art or often uses an specific type of censorship, all of those elements will carry on to the output even with the respective negatives and sometimes even with positive tags like "uncensored", which should be a bit stronger than the negatives. The only way to get rid of things like that seems to be lowering the artist strength or just training a lora for the style/concept manually with clean images. Any plans to fix this or will it stay for the full release?
Are there any plans to make styles more consistent? From my generations, I've noticed styles can change the output from different seeds with the same prompt, dramatically, especially when mixing multiple artist tags. Backgrounds on the other hand, keep a similar shape between different seeds, the biggest thing changing being the style.
Can you adjust the strength the same way you do in most sdxl models? Like (vanishing point:1.2) or is there a different way to adjust strength or what would be the equivalent strength to achieve a similar one to the one you get with the usual clip text encoders on normal sdxl models with comfy's or auto1111's weight interpretation? Because it seems to have a weaker effect, not sure if it's from the qwen text encoder not communicating properly or something else.
Can you adjust the strength the same way you do in most sdxl models?
https://huggingface.co/circlestone-labs/Anima/discussions/135#69e9e6fcead129b9d5b8be59
