Style of MiyuAI for Anima model.

I put a lot of time and effort to make this 2nd version be as good as possible.

These tags help produce somewhat better results:

(masterpiece, highly detailed, realistic:1.5), vivid colors, weight shifted, (dutch angle:0.5), (depth of field:0.5),
(foreshortening:2)

(foreshortening tag makes it look weird in some specific cases, but it helps a lot in most others).

This lora works much better with high image resolutions 1536x1536 and variations of this size, like 1248x1872 for 2:3).

I also recommend experimenting with samplers and schedulers. I haven't seen anyone use rk sampler and it's doing wonders for me, and it's really fast as well (all images done using turbo lora + cosmos predict at 0.2 weight, based on base model). In my opinion increasing image size (up to a certain point) improves detail much more than increasing steps.

Also check out Anima Booster nodes, they introduce a massive speedup in generation (workflow embedded on these images uses them).