Koushui Style v2 for **Anima** — fine clean lineart, flat high-saturation fills, warm colour temperature and a soft glow on skin, with carefully detailed faces. This is a **pure style LoRA**: no character, no outfit. v2 is a full retrain on a much larger, properly tagged dataset (94 curated illustrations, auto-tagged and caption-cleaned), replacing the v1 release (40 images, rank 16). Rank was raised to 32 and dataset repeats tuned so the style reads correctly at low weight without repainting the whole image. ## How to use ``` koushui, 1girl, ``` Use weight **0.6**: at 1.0 colours blow out, at 0.6 you keep Anima's anatomy while adopting the palette and line feel. **Known-good stack:** `koushui` 0.6 + `chaozhi` 0.8. ## Recommended settings | Item | Value | |-|-| | Base model | Anima (aesthetic v1.x) | | LoRA weight | 0.6 | | Steps | 25–35 | | CFG | 5–7 | ## Known limitations - Strongly pushes warm/saturated palettes — cold scenes (snow, night, blue hour) need explicit colour cues or weight 0.4. - Stacks with any character, but expect that character's hair/eye colours to shift warm (compensate in the prompt). ## Training details 94 curated images · rank 32 / alpha 32 · LR 1e-4 · 5 repeats · caption shuffle · dropout 0.15 · 5 epochs (~2h on a 3060 Ti) · costumes/characters disabled via caption cleaning. Style LoRA trained for personal use. If you are the original artist and want this removed, contact me and I will take it down.