Penny seems fine with sidelocks in negative. Atsukos outfit maybe needs more training. The shirt is not coming out very stable. Putting "patched jeans" in the negative fixed the jeans for her in solo pics. Also not using a pose or framing in the prompt always gives me sideview, maybe because of "wristwatch"?
For prompting style with anima you need to use "@" in front of the style tag to get the full impact, e.g. "@1980 \(style\)" if you used "1980 (style)" in the caption (I am using \ to escape the bracket, don't know if this works on the civitai generator, but that's how brackets in tags should be handled in Forge).
I do not know how it should look, but it looks different like that. Also you can try increasing the weight of a tag pretty much with anima, e.g. use :2 instead of :1.2.

Edit: 2 of the 6 images I posted to the model page are flagged for review... sigh

@NanashiAnon No problem. I have a few suggestions regarding the style.
I think the main difference between your illustrious loras and this one is that your illustrious loras have text encoder layers. The obvious solution for me would be to train the text encoder. The anima model I trained that is (in terms of style) closest to your loras is my Bibi Blocksberg lora: https://civitai.red/models/2586297/bibi-blocksberg-little-witch-anima?modelVersionId=2905437

It's partially trained with cartoon style images from late 90s cartoon series (+ additional artwork). I even used the tags "1980s (style)" and "toon (style)" in the caption of the cartoon screencaps. Here are some examples using the lora, unfortunately two of the four images were tagged falsely with underwear and are flagged for review, so I am giving you the direct links:
without style tag
using "toon \(style\), 1980s \(style\)" in the prompt
using only "1980s \(style\)" in the prompt
using "cartoon" in the prompt
Using both style tags I used in training caption gives a rather good approximation of the cartoon style I think (there's a bit of ai filter on it of course), here is a screencap of the series:
https://i.pinimg.com/originals/c2/db/12/c2db12c7826f3e5877b99d59fb33e1e8.png

Since text encoder training is currently not available (I don't think it will fix any problem with anima training, but maybe you can see where I am coming from saying this should be the first thing done in improving anima training on civitai), I have another suggestion.

Since "1980s (style)" is a danbooru tag, the images associated with it in training the base model are mostly anime images. I guess doing text encoder training you can overwrite the style pretty good, but it does not work to the same degree without it (maybe with far longer training). So I suggest until text encoder training gets available you should try to use a different tag for the style. Here is a lora trained by Citron without text encoder layers:
https://civitai.red/models/1643137/king-k-rool-or-donkey-kong-il-and-anima-and-zit

It's different from your loras of course, but there were cartoon images used in the training data and Citron reports that using "flat color" and "cartoon" as caption for the style does not produce a generic cartoon style, but reproduces the style in the training data (see my comment and the answer on the model page). This was trained with preview1 version of anima though, so I do not know if the same applies for anima base.

Another suggestion, maybe use a custom trigger word for the style. You have a two character LoRA, so that should work in principle, also you could add additional images with the same style but different characters to the dataset and tag all of them with the custom style trigger.
Other than that just training for more steps will probably help, too.

Sorry for writing a wall of text :D